Published first on TheFly – the ultimate source for real-time, market-moving breaking financial news.

Cigna revives merger talks with Humana, Bloomberg reports

Shares of Humana, which has a market capitalization of about $32 billion, were up about 6% in after-hours trading on Friday, while those of Cigna were down about 5%. Cigna was valued at about $94 billion, according to data compiled by LSEG. Cigna and Humana declined to comment.

Cigna resumes mega merger talks with rival health insurer Humana, Bloomberg News reports

Shares of Humana were up about 5% in after-hours trading on Friday, while those of Cigna were down about 2%. Last year, Reuters reported that Cigna ended its attempt to negotiate an acquisition of Humana after the pair failed to agree on a price. Cigna did not immediately respond to a Reuters ...

Cigna resumes merger discussions with Humana, Bloomberg News reports

(Bloomberg) — Humana Inc. (HUM) investors haven’t struggled through anything this bad since the global financial crisis 15 years ago.

Medicare mess sends Humana shares on worst fall since financial crisis

and excellence with plans to fit every budget, wellness goal, and lifestyle. Depending on the chosen plan, physical and mental health benefits include fitness allowances, pediatric dental care, mental health treatment, and even virtual health and urgent ...

Brevard County-based Health First Health Plans Expands Three Decades of Experience and Service to Central Florida

Advanta Health Solutions, Inc. ( a leading health technology company that helps plans and plan sponsors improve the physical and mental wellbeing of their employees, members, and customers, announced today a new technology partnership with ROOK ( an API for health wearable data.

Advanta Health Partners with ROOK to Expand Biomarker Insights and Data-Driven Capabilities within its ActiveFit+ Platform

The company is working on big things, but employees disagree over whether they should be serving people who are healthy or sick.

Apple has plans to eventually, maybe revolutionize health care

The $81 million in federal funding was given to Rhode Island earlier this year. Now municipalities are thinking of how to spend it.

RI giving out millions to cities and towns for 'learning centers.' Here's how they would use it.

It’s 6 a.m. Some 25 associates who work on the battery production floor at Foxconn’s Ohio plant in Lordstown gather for a morning ritual. These Foxconn employees aren’t in store for an intense pep talk or

At Foxconn Ohio, Fitness and Health Are Priorities

Wellness Coach offers a well-rounded approach to tracking and improving your physical, mental, and sleep health. Here's why we love it.

7 Things We Love About the Wellness Coach App

Humana

humana

New Framework Simplifies Comparison of Language Processing Tools Across Multiple Languages

Yeni Çərçivə Dil Emalı Alətlərinin Çox Dillər üzrə Müqayisəsini Sadələşdirir

Новая структура спрашчае параўнанне інструментаў апрацоўкі мовы на некалькіх мовах

Novi okvir pojednostavljuje poređenje alata za obradu jezika na više jezika

Un nuevo marco simplifica la comparación de herramientas de procesamiento del lenguaje en varios idiomas

नया फ्रेमवर्क विभिन्न भाषाओं में भाषा प्रसंस्करण उपकरणों की तुलना को सरल बनाता है

Նոր շրջանակը պարզեցնում է լեզուների մշակման գործիքների համեմատությունը բազմաթիվ լեզուներով

Nuovo framework semplifica il confronto degli strumenti di elaborazione del linguaggio in più lingue

新しいフレームワークにより、複数の言語にわたる言語処理ツールの比較が簡素化されます

Rangka Kerja Baharu Memudahkan Perbandingan Alat Pemprosesan Bahasa Merentas Pelbagai Bahasa

Nový rámec zjednodušuje porovnávanie nástrojov na spracovanie jazykov vo viacerých jazykoch

Mfumo Mpya Hurahisisha Ulinganisho wa Zana za Kuchakata Lugha Katika Lugha Nyingi

Uhlaka Olusha Lenza Lula Ukuqhathanisa Amathuluzi Okucubungula Ulimi Kuzo Zonke Izilimi Eziningi

Every AI Prediction That Came True in 2024—and the Ones That Didn't

Todas las predicciones de IA que se cumplieron en 2024 (y las que no)

Toutes les prédictions de l'IA qui se sont réalisées en 2024, et celles qui ne se sont pas réalisées

Todas as predicións da IA que se fixeron realidade en 2024 e as que non o fixeron

2024-ին իրականություն դարձած յուրաքանչյուր AI կանխատեսում, և նրանք, որոնք չեն իրականացել

כל תחזית AI שהתגשמה בשנת 2024 - ואלה שלא

2024 年に実現したすべての AI 予測と実現しなかった予測

ხელოვნური ინტელექტის ყველა პროგნოზი, რომელიც ახდა 2024 წელს და ის, რაც არ შესრულდა

រាល់ការទស្សន៍ទាយ AI ដែលបានក្លាយជាការពិតនៅឆ្នាំ 2024 ហើយអ្វីដែលមិនបានកើតឡើង

2024년에 실현된 모든 AI 예측과 실현되지 않은 예측

Wszystkie prognozy dotyczące sztucznej inteligencji, które spełniły się w 2024 r. — i te, które się nie spełniły

Todas as previsões de IA que se concretizaram em 2024 — e as que não se concretizaram

2024 දී සැබෑ වූ සෑම AI අනාවැකියක්ම - සහ නොකළ ඒවා

The Sneaky Way Web Browsers Are Identifying You (Even When You Turn Off Cookies)

Die skelm manier waarop webblaaiers jou identifiseer (selfs wanneer jy koekies afskakel)

Вэб-браўзеры ідэнтыфікуюць вас (нават калі вы адключаеце файлы cookie)

Ο ύπουλος τρόπος Τα προγράμματα περιήγησης Ιστού σας αναγνωρίζουν (ακόμα και όταν απενεργοποιείτε τα cookies)

La forma encubierta en que los navegadores web te identifican (incluso cuando desactivas las cookies)

वेब ब्राउज़र आपको गुप्त तरीके से पहचानते हैं (तब भी जब आप कुकीज़ बंद कर देते हैं)

Fason sournwa navigatè entènèt yo idantifye ou (menm lè ou fèmen bonbon)

The Sneaky Way վեբ զննարկիչները նույնականացնում են ձեզ (նույնիսկ երբ անջատում եք թխուկները)

ウェブブラウザがあなたを識別する巧妙な方法（クッキーをオフにしている場合でも）

The Sneaky Way ვებ ბრაუზერები იდენტიფიცირებენ თქვენ (მაშინაც კი, როდესაც გამორთავთ ქუქი-ფაილებს)

द स्नीकी वे वेब ब्राउजरहरूले तपाईंलाई पहिचान गर्दैछन् (तपाईंले कुकीहरू बन्द गर्दा पनि)

Mënyra e poshtër Shfletuesit e uebit po ju identifikojnë (edhe kur i çaktivizoni cookies)

网络浏览器以狡猾的方式识别你（即使你关闭了 Cookie）

How Effective Are AI Models at Biomedical Problem Solving?

Table of Links Abstract and 1. Introduction Background and Related Work


Threat Model


Robust Style Mimicry


Experimental Setup


Results
6.1 Main Findings: All Protections are Easily Circumvented
6.2 Analysis


Discussion and Broader Impact, Acknowledgements, and References A. Detailed Art Examples B. Robust Mimicry Generations C. Detailed Results D. Differences with Glaze Finetuning E. Findings on Glaze 2.0 F. Findings on Mist v2 G. Methods for Style Mimicry H. Existing Style Mimicry Protections I. Robust Mimicry Methods J. Experimental Setup K. User Study L. Compute Resources 5 Experimental Setup Protection tools. We evaluate three protection tools—Mist, Glaze and Anti-DreamBooth—against four robust mimicry methods—Gaussian noising, DiffPure, Noisy Upscaling and IMPRESS++—and a baseline mimicry method. We refer to a combination of a protection tool and a mimicry method as a scenario. We thus analyze fifteen possible scenarios. Appendix J describes our experimental setup for style mimicry and protections in detail. Artists. We evaluate each style mimicry scenario on images from 10 different artists, which we selected to maximize style diversity. To address limitations in prior evaluations (Shan et al., 2023b), we use five historical artists as well as five contemporary artists who are unlikely to be highly represented in the generative model’s training set (two of these were also used in Glaze’s evaluation).[2] All details about artist selection are included in Appendix J. Implementation. Our mimicry methods finetune Stable Diffusion 2.1 (Rombach et al., 2022), the best open-source model available at the time when the protections we study were introduced. We use an off-the-shelf finetuning script from HuggingFace (see Appendix J.1 for details). We first validate that our style mimicry pipeline is successful on unprotected art using a user study, detailed in Appendix K.1. For protections, we use the original codebases to reproduce Mist and Anti-Dreambooth. Since Glaze does not have a public codebase (and the authors were unable to share one), we use the released Windows application binary (version 1.1.1) as a black-box. We set each scheme’s hyperparameters to maximize protections. See Appendix J.2 for details on the configuration for each protection. We perform robust mimicry by finetuning on 18 different images per artist. We then generate images for 10 different prompts. These prompts are designed to cover diverse motifs that the base model, Stable Diffusion 2.1, can successfully generate. See Appendix K for details about prompt design. User study. To measure the success of each style mimicry scenario, we rely only on human evaluations since previous work found automated metrics (e.g., using CLIP (Radford et al., 2021)) to be unreliable (Shan et al., 2023a,b). Moreover, style protections not only prevent style transfer, but also reduce the overall quality of the generated images (see Figure 3 for examples). We thus design a user study to evaluate image quality and style transfer as independent attributes of the generations.[3] Our user study asks annotators on Amazon Mechanical Turk (MTurk) to compare image pairs, where one image is generated by a robust mimicry method, and the other from a baseline state-of-the-art mimicry method that uses unprotected art of the artist. A perfectly robust mimicry method would generate images of quality and style indistinguishable from those generated directly from unprotected art. We perform two separate studies: one assessing image quality (e.g., which image looks “better”) and another evaluating stylistic transfer (i.e., which image captures the artist’s original style better, disregarding potential quality artifacts). Our results show that these two metrics obtain very similar results across all scenarios. Appendix K describes our user study and interface in detail. As noted by the authors of Glaze (Shan et al., 2023a), the users of platforms like MTurk might not have high artistic expertise. However, we believe that the judgment of non-artists is also relevant as they ultimately represent a large fraction of the consumers of digital art. Thus, if lay people consider mimicry attempts to be successful, mimicked art could hurt an artist’s business. Also, to mitigate potential issues with the quality of annotations (Kennedy et al., 2020), we put in place several control mechanisms to filter out low-quality annotations to the best of our abilities (details in Appendix K). Evaluation metric. We define the success rate of a robust mimicry method as the percentage of annotators (5 per comparison) who prefer outputs from the robust mimicry method over those from a baseline method finetuned on unprotected art (when judging either style match or overall image quality). Formally, we define the success rate for an artist in a specific scenario as: A perfectly robust mimicry method would thus obtain a success rate of 50%, indicating that its outputs are indistinguishable in quality and style from those from the baseline, unprotected method. In contrast, a very successful protection would result in success rates of around 0% for robust mimicry methods, indicating that mimicry on top of protected images always yields worse outputs. Authors:
(1) Robert Honig, ETH Zurich (robert.hoenig@inf.ethz.ch);
(2) Javier Rando, ETH Zurich (javier.rando@inf.ethz.ch);
(3) Nicholas Carlini, Google DeepMind;
(4) Florian Tramer, ETH Zurich (florian.tramer@inf.ethz.ch). This paper is available on arxiv under CC BY 4.0 license. [2] Contemporary Artists were selected from Artstation. We keep them anonymous throughout this work—and refrain from showcasing their art—except for artists who gave us explicit permission to share their identity and art. We will share all images used in our experiments upon request with researchers. [3] The user study was approved by our institution’s IRB. Table of Links Abstract and 1. Introduction Abstract and 1. Introduction Background and Related Work Threat Model Robust Style Mimicry Experimental Setup Results
6.1 Main Findings: All Protections are Easily Circumvented
6.2 Analysis Discussion and Broader Impact, Acknowledgements, and References Background and Related Work Background and Related Work Background and Related Work Threat Model Threat Model Threat Model Robust Style Mimicry Robust Style Mimicry Robust Style Mimicry Experimental Setup Experimental Setup Experimental Setup Results 6.1 Main Findings: All Protections are Easily Circumvented 6.2 Analysis Results Results 6.1 Main Findings: All Protections are Easily Circumvented 6.1 Main Findings: All Protections are Easily Circumvented 6.2 Analysis 6.2 Analysis Discussion and Broader Impact, Acknowledgements, and References Discussion and Broader Impact, Acknowledgements, and References Discussion and Broader Impact, Acknowledgements, and References A. Detailed Art Examples A. Detailed Art Examples B. Robust Mimicry Generations B. Robust Mimicry Generations C. Detailed Results C. Detailed Results D. Differences with Glaze Finetuning D. Differences with Glaze Finetuning E. Findings on Glaze 2.0 E. Findings on Glaze 2.0 F. Findings on Mist v2 F. Findings on Mist v2 G. Methods for Style Mimicry G. Methods for Style Mimicry H. Existing Style Mimicry Protections H. Existing Style Mimicry Protections I. Robust Mimicry Methods I. Robust Mimicry Methods J. Experimental Setup J. Experimental Setup K. User Study K. User Study L. Compute Resources L. Compute Resources 5 Experimental Setup Protection tools . We evaluate three protection tools—Mist, Glaze and Anti-DreamBooth—against four robust mimicry methods—Gaussian noising, DiffPure, Noisy Upscaling and IMPRESS++—and a baseline mimicry method. We refer to a combination of a protection tool and a mimicry method as a scenario. We thus analyze fifteen possible scenarios. Appendix J describes our experimental setup for style mimicry and protections in detail. Protection tools Artists . We evaluate each style mimicry scenario on images from 10 different artists, which we selected to maximize style diversity. To address limitations in prior evaluations (Shan et al., 2023b), we use five historical artists as well as five contemporary artists who are unlikely to be highly represented in the generative model’s training set (two of these were also used in Glaze’s evaluation).[2] All details about artist selection are included in Appendix J. Artists Implementation . Our mimicry methods finetune Stable Diffusion 2.1 (Rombach et al., 2022), the best open-source model available at the time when the protections we study were introduced. We use an off-the-shelf finetuning script from HuggingFace (see Appendix J.1 for details). We first validate that our style mimicry pipeline is successful on unprotected art using a user study, detailed in Appendix K.1. For protections, we use the original codebases to reproduce Mist and Anti-Dreambooth. Since Glaze does not have a public codebase (and the authors were unable to share one), we use the released Windows application binary (version 1.1.1) as a black-box. We set each scheme’s hyperparameters to maximize protections. See Appendix J.2 for details on the configuration for each protection. Implementation We perform robust mimicry by finetuning on 18 different images per artist. We then generate images for 10 different prompts. These prompts are designed to cover diverse motifs that the base model, Stable Diffusion 2.1, can successfully generate. See Appendix K for details about prompt design. User study. To measure the success of each style mimicry scenario, we rely only on human evaluations since previous work found automated metrics (e.g., using CLIP (Radford et al., 2021)) to be unreliable (Shan et al., 2023a,b). Moreover, style protections not only prevent style transfer, but also reduce the overall quality of the generated images (see Figure 3 for examples). We thus design a user study to evaluate image quality and style transfer as independent attributes of the generations.[3] User study. Our user study asks annotators on Amazon Mechanical Turk (MTurk) to compare image pairs, where one image is generated by a robust mimicry method, and the other from a baseline state-of-the-art mimicry method that uses unprotected art of the artist. A perfectly robust mimicry method would generate images of quality and style indistinguishable from those generated directly from unprotected art. We perform two separate studies: one assessing image quality (e.g., which image looks “better”) and another evaluating stylistic transfer (i.e., which image captures the artist’s original style better, disregarding potential quality artifacts). Our results show that these two metrics obtain very similar results across all scenarios. Appendix K describes our user study and interface in detail. As noted by the authors of Glaze (Shan et al., 2023a), the users of platforms like MTurk might not have high artistic expertise. However, we believe that the judgment of non-artists is also relevant as they ultimately represent a large fraction of the consumers of digital art. Thus, if lay people consider mimicry attempts to be successful, mimicked art could hurt an artist’s business. Also, to mitigate potential issues with the quality of annotations (Kennedy et al., 2020), we put in place several control mechanisms to filter out low-quality annotations to the best of our abilities (details in Appendix K). Evaluation metric. We define the success rate of a robust mimicry method as the percentage of annotators (5 per comparison) who prefer outputs from the robust mimicry method over those from a baseline method finetuned on unprotected art (when judging either style match or overall image quality). Formally, we define the success rate for an artist in a specific scenario as: Evaluation metric. A perfectly robust mimicry method would thus obtain a success rate of 50%, indicating that its outputs are indistinguishable in quality and style from those from the baseline, unprotected method. In contrast, a very successful protection would result in success rates of around 0% for robust mimicry methods, indicating that mimicry on top of protected images always yields worse outputs. Authors: (1) Robert Honig, ETH Zurich (robert.hoenig@inf.ethz.ch); (2) Javier Rando, ETH Zurich (javier.rando@inf.ethz.ch); (3) Nicholas Carlini, Google DeepMind; (4) Florian Tramer, ETH Zurich (florian.tramer@inf.ethz.ch). Authors: Authors: (1) Robert Honig, ETH Zurich (robert.hoenig@inf.ethz.ch); (2) Javier Rando, ETH Zurich (javier.rando@inf.ethz.ch); (3) Nicholas Carlini, Google DeepMind; (4) Florian Tramer, ETH Zurich (florian.tramer@inf.ethz.ch). This paper is available on arxiv under CC BY 4.0 license. This paper is available on arxiv under CC BY 4.0 license. available on arxiv available on arxiv [2] Contemporary Artists were selected from Artstation. We keep them anonymous throughout this work—and refrain from showcasing their art—except for artists who gave us explicit permission to share their identity and art. We will share all images used in our experiments upon request with researchers. [3] The user study was approved by our institution’s IRB.

Can AI Mimic Famous Art Styles Despite Protective Measures?

Technological progress, especially in artificial intelligence and robotics, is rapidly making human workers less essential. Companies are shifting to robots and AI because they are cheaper, more efficient, and never need breaks or benefits. In many industries, automation is replacing people entirely, and this trend is only accelerating. As machines handle more tasks better and faster than humans, the need for human labor is shrinking. This creates serious challenges: tax systems that depend on workers' incomes are breaking down, inequality is growing, and social safety nets are at risk. If this continues, we face a future where people may no longer be needed in many jobs, leaving society to figure out how to redefine work, purpose, and economic survival. According to a McKinsey study, around 45% of all workplace tasks can already be automated today. This makes automation a key strategy for reducing business costs, particularly in sectors such as manufacturing, logistics, and even knowledge-based work. These trends raise critical questions about how to adapt tax systems and income redistribution in a world where the core economic activity is increasingly driven by machines rather than humans. Why Are Robots and AI Cheaper Than Humans? Robots and AI provide businesses with significant advantages, making them far more cost-effective than human workers. First, robots have much lower operating costs. Once the initial expense of development and setup is covered, maintenance costs are minimal. They don’t need salaries, vacations, or sick days, making them an ideal long-term investment. Second, robots outperform humans in productivity and efficiency. They can operate 24/7 without fatigue or quality drops, which is a game-changer for industries like manufacturing, where automation can boost output by 40–50%. For instance, companies like Amazon already use robots in their warehouses to speed up processes and reduce costs. The greatest advantage for businesses, however, is that robots free them from social responsibilities. There’s no need for pensions, health insurance, or paid time off. On top of that, robots and AI are incredibly adaptable—new tasks can be handled through simple software updates, eliminating the need to retrain employees. This flexibility, combined with cost savings, makes automation especially attractive in competitive industries where efficiency and cost-cutting are critical. The Impact of Automation on the Economy and Taxation in the Near Future The rise of automation is set to plunge traditional tax systems into an unavoidable crisis. As human labor declines, so does the pool of taxable income, leaving government budgets increasingly strained—especially in nations like Norway, Sweden, and Denmark, where income taxes form a substantial portion of public revenue. Meanwhile, the economic gains from automation are becoming concentrated in the hands of technology owners and intellectual property holders, deepening the divide between the wealthy and everyone else. The World Economic Forum warns that without innovative approaches to taxation, this gap will only widen. Progressive taxation, once a cornerstone of income redistribution, is losing its effectiveness in a world where technology, not labor, drives wealth creation. Tax systems designed for an industrial economy can no longer address the challenges of a technology-driven one. To ensure a fairer distribution of wealth and maintain fiscal stability, societies must embrace bold, forward-thinking taxation models that align with the realities of our evolving economy. Possible Approaches to Reforming Tax Systems One potential solution to address the challenges of automation is to introduce a tax on robots. Bill Gates has suggested that companies adopting automation should pay a tax equivalent to the income taxes of the workers they replace. The revenue from this tax could support social programs or fund worker retraining. However, this idea remains controversial, as it could slow the pace of innovation—a prospect that, as a technological optimist, I find concerning. A more balanced alternative might involve hybrid models that combine a robot tax with other forms of taxation. Another promising avenue is taxing data and the digital economy. Companies like Google and Meta generate enormous profits from user-generated data, yet contribute disproportionately little to public revenue. The European Union is already exploring the introduction of a digital tax to address the economic disparities created by these tech giants' dominance. Sam Altman, CEO of OpenAI, has proposed shifting from taxing labor to taxing capital in the context of the digital economy. This could include taxes on profits generated by automation or on capital investments tied to AI. Such measures could help redistribute the wealth created by technology and fund essential social programs. A further transformative approach is the implementation of Universal Basic Income (UBI), which provides regular, unconditional payments to all citizens. Trials in Finland and other countries have shown that UBI can reduce poverty, encourage entrepreneurship, and enhance overall quality of life. This concept could serve as a safety net in an economy where machines increasingly replace human labor. As AI and robotics continue to make automation more economically advantageous than human labor, the economic landscape is undergoing a profound transformation. Traditional taxation systems, particularly progressive income taxes, are becoming outdated in an era where wealth is generated more by machines and software than by people. These systems are no longer sufficient to sustain social equity or maintain government budgets. The future of taxation must align with the realities of an automated economy. Solutions like taxing robots, data, and capital, along with introducing Universal Basic Income, offer a roadmap to adapt. Such measures can help offset the decline in traditional tax revenues, reduce economic inequality, and ensure economic stability. Governments must act now to modernize their tax systems and prepare for the profound changes brought about by the age of automation. Technological progress, especially in artificial intelligence and robotics, is rapidly making human workers less essential. Companies are shifting to robots and AI because they are cheaper, more efficient, and never need breaks or benefits. In many industries, automation is replacing people entirely, and this trend is only accelerating. As machines handle more tasks better and faster than humans, the need for human labor is shrinking. This creates serious challenges: tax systems that depend on workers' incomes are breaking down, inequality is growing, and social safety nets are at risk. If this continues, we face a future where people may no longer be needed in many jobs, leaving society to figure out how to redefine work, purpose, and economic survival. According to a McKinsey study , around 45% of all workplace tasks can already be automated today. This makes automation a key strategy for reducing business costs, particularly in sectors such as manufacturing, logistics, and even knowledge-based work. These trends raise critical questions about how to adapt tax systems and income redistribution in a world where the core economic activity is increasingly driven by machines rather than humans. McKinsey study McKinsey study Why Are Robots and AI Cheaper Than Humans? Robots and AI provide businesses with significant advantages, making them far more cost-effective than human workers. First, robots have much lower operating costs. Once the initial expense of development and setup is covered, maintenance costs are minimal. They don’t need salaries, vacations, or sick days, making them an ideal long-term investment. Second, robots outperform humans in productivity and efficiency. They can operate 24/7 without fatigue or quality drops, which is a game-changer for industries like manufacturing, where automation can boost output by 40–50% . For instance, companies like Amazon already use robots in their warehouses to speed up processes and reduce costs. can boost output by 40–50% can boost output by 40–50% Amazon already use robots Amazon already use robots The greatest advantage for businesses, however, is that robots free them from social responsibilities. There’s no need for pensions, health insurance, or paid time off. On top of that, robots and AI are incredibly adaptable—new tasks can be handled through simple software updates, eliminating the need to retrain employees. This flexibility, combined with cost savings, makes automation especially attractive in competitive industries where efficiency and cost-cutting are critical. The Impact of Automation on the Economy and Taxation in the Near Future The rise of automation is set to plunge traditional tax systems into an unavoidable crisis. As human labor declines, so does the pool of taxable income, leaving government budgets increasingly strained—especially in nations like Norway, Sweden, and Denmark, where income taxes form a substantial portion of public revenue. Meanwhile, the economic gains from automation are becoming concentrated in the hands of technology owners and intellectual property holders, deepening the divide between the wealthy and everyone else. The World Economic Forum warns that without innovative approaches to taxation, this gap will only widen. The World Economic Forum warns The World Economic Forum warns Progressive taxation, once a cornerstone of income redistribution, is losing its effectiveness in a world where technology, not labor, drives wealth creation. Tax systems designed for an industrial economy can no longer address the challenges of a technology-driven one. To ensure a fairer distribution of wealth and maintain fiscal stability, societies must embrace bold, forward-thinking taxation models that align with the realities of our evolving economy. Possible Approaches to Reforming Tax Systems One potential solution to address the challenges of automation is to introduce a tax on robots. Bill Gates has suggested that companies adopting automation should pay a tax equivalent to the income taxes of the workers they replace. The revenue from this tax could support social programs or fund worker retraining. However, this idea remains controversial, as it could slow the pace of innovation—a prospect that, as a technological optimist, I find concerning. A more balanced alternative might involve hybrid models that combine a robot tax with other forms of taxation. Bill Gates has suggested Bill Gates has suggested Another promising avenue is taxing data and the digital economy. Companies like Google and Meta generate enormous profits from user-generated data, yet contribute disproportionately little to public revenue. The European Union is already exploring the introduction of a digital tax to address the economic disparities created by these tech giants' dominance. introduction of a digital tax introduction of a digital tax Sam Altman, CEO of OpenAI, has proposed shifting from taxing labor to taxing capital in the context of the digital economy. This could include taxes on profits generated by automation or on capital investments tied to AI. Such measures could help redistribute the wealth created by technology and fund essential social programs. has proposed has proposed A further transformative approach is the implementation of Universal Basic Income (UBI), which provides regular, unconditional payments to all citizens. Trials in Finland and other countries have shown that UBI can reduce poverty, encourage entrepreneurship, and enhance overall quality of life. This concept could serve as a safety net in an economy where machines increasingly replace human labor. Trials in Finland Trials in Finland As AI and robotics continue to make automation more economically advantageous than human labor, the economic landscape is undergoing a profound transformation. Traditional taxation systems, particularly progressive income taxes, are becoming outdated in an era where wealth is generated more by machines and software than by people. These systems are no longer sufficient to sustain social equity or maintain government budgets. The future of taxation must align with the realities of an automated economy. Solutions like taxing robots, data, and capital, along with introducing Universal Basic Income, offer a roadmap to adapt. Such measures can help offset the decline in traditional tax revenues, reduce economic inequality, and ensure economic stability. Governments must act now to modernize their tax systems and prepare for the profound changes brought about by the age of automation.

Experts Aren't Discussing Nearly Enough the Tax Implications of Replacing Workers With AI 

ባለሙያዎች ሠራተኞችን በአይአይ መተካት የሚያስከትለውን የግብር አንድምታ በበቂ ሁኔታ እየተወያዩ አይደሉም። 

Mütəxəssislər işçilərin süni intellektlə əvəzlənməsinin vergi təsirlərini demək olar ki, kifayət qədər müzakirə etmirlər 

Los expertos no discuten lo suficiente las implicaciones fiscales de reemplazar trabajadores por IA

Adituek ez dute ia nahikoa eztabaidatzen langileak AIrekin ordezkatzearen zerga-ondorioez 

Les experts n’abordent pas suffisamment les implications fiscales du remplacement des travailleurs par l’IA

विशेषज्ञ कर्मचारियों को एआई से बदलने के कर निहितार्थों पर पर्याप्त चर्चा नहीं कर रहे हैं

専門家は労働者をAIに置き換えることによる税務上の影響について十分に議論していない

विशेषज्ञहरूले एआईको साथ कामदारहरू प्रतिस्थापन गर्ने कर प्रभावहरू लगभग पर्याप्त छलफल गरिरहेका छैनन् 

Ditsebi Ga di Boledišane Mo e nyakilego go ba ka mo go lekanego ka Ditlamorago tša Motšhelo tša go Tšeela Bašomi Legato ka AI 

Стручњаци не разговарају ни изблиза довољно о пореским импликацијама замене радника вештачком интелигенцијом 

Halos Hindi Sapat na Tinatalakay ng mga Eksperto ang mga Implikasyon sa Buwis ng Pagpapalit sa mga Manggagawa ng AI 

Uzmanlar, İşçilerin Yapay Zeka ile Değiştirilmesinin Vergisel Etkilerini Yeterince Tartışmıyor

Table of Links Abstract and Introduction
Related Work
Problem Definition
Method
Experiments
Conclusion and References A. Appendix A.1. Full Prompts and A.2 ICPL Details A. 3 Baseline Details A.4 Environment Details A.5 Proxy Human Preference A.6 Human-in-the-Loop Preference 2 RELATED WORK Reward Design. In reinforcement learning, reward design is a core challenge, as rewards must both represent a desired set of behaviors and provide enough signal for learning. The most common approach to reward design is handcrafting, which requires a large number of trials by experts (Sutton, 2018; Singh et al., 2009). Since hand-coded reward design requires extensive engineering effort, several prior works have studied modeling the reward function with precollected data. For example, Inverse Reinforcement Learning (IRL) aims to recover a reward function from expert demonstration data (Arora & Doshi, 2021; Ng et al., 2000). With advances in pretrained foundation models, some recent works have also studied using large language models or vision-language models to provide reward signals (Ma et al., 2022; Fan et al., 2022; Du et al., 2023; Karamcheti et al., 2023; Kwon et al., 2023; Wang et al., 2024; Ma et al., 2024). Among these approaches, EUREKA (Ma et al., 2023) is the closest to our work, instructing the LLM to generate and select novel reward functions based on environment feedback with an evolutionary framework. However, as opposed to performing preference learning, EUREKA uses the LLM to design dense rewards that help with learning a hand-designed sparse reward. In contrast, ICPL focuses on finding a reward that correctly orders a set of preferences. However, we note that EUREKA also has a small, preliminary investigation combining human preferences with an LLM to generate human-preferred behaviors in a single scenario. We note that this experiment had the human provide their feedback in text, whereas we only use preference queries. This paper is a significantly scaled-up version of that investigation as well as a methodological study of how best to incorporate prior rounds of feedback. Human-in-the-loop Reinforcement Learning. Feedback from humans has been proven to be effective in training reinforcement learning agents that better match human preferences (Retzlaff et al., 2024; Mosqueira-Rey et al., 2023; Kwon et al., 2023). Previous works have investigated human feedback in various forms, such as trajectory comparisons, preferences, demonstrations, and corrections (Wirth et al., 2017; Ng et al., 2000; Jeon et al., 2020; Peng et al., 2024). Among these various methods, preference-based RL has been successfully scaled to train large foundation models for hard tasks like dialogue, e.g. ChatGPT (Ouyang et al., 2022). In LLM-based applications, prompting is a simple way to provide human feedback in order to align LLMs with human preferences (Giray, 2023; White et al., 2023; Chen et al., 2023). Iteratively refining the prompts with feedback from the environment or human users has shown promise in improving the output of the LLM (Wu et al., 2021; Nasiriany et al., 2024). This work extends the usage of the ability to control LLM behavior via in-context prompts. We aim to utilize interactive rounds of preference feedback between the LLM and humans to guide the LLM to generate reward functions that can elicit behaviors that align with human preferences. Authors:
(1) Chao Yu, Tsinghua University;
(2) Hong Lu, Tsinghua University;
(3) Jiaxuan Gao, Tsinghua University;
(4) Qixin Tan, Tsinghua University;
(5) Xinting Yang, Tsinghua University;
(6) Yu Wang, with equal advising from Tsinghua University;
(7) Yi Wu, with equal advising from Tsinghua University and the Shanghai Qi Zhi Institute;
(8) Eugene Vinitsky, with equal advising from New York University (zoeyuchao@gmail.com). This paper is available on arxiv under CC 4.0 license. Table of Links Abstract and Introduction Related Work Problem Definition Method Experiments Conclusion and References Abstract and Introduction Abstract and Introduction Related Work Related Work Problem Definition Problem Definition Method Method Experiments Experiments Conclusion and References Conclusion and References A. Appendix A. Appendix A.1. Full Prompts and A.2 ICPL Details A.1. Full Prompts and A.2 ICPL Details A. 3 Baseline Details A. 3 Baseline Details A.4 Environment Details A.4 Environment Details A.5 Proxy Human Preference A.5 Proxy Human Preference A.6 Human-in-the-Loop Preference A.6 Human-in-the-Loop Preference 2 RELATED WORK Reward Design . In reinforcement learning, reward design is a core challenge, as rewards must both represent a desired set of behaviors and provide enough signal for learning. The most common approach to reward design is handcrafting, which requires a large number of trials by experts (Sutton, 2018; Singh et al., 2009). Since hand-coded reward design requires extensive engineering effort, several prior works have studied modeling the reward function with precollected data. For example, Inverse Reinforcement Learning (IRL) aims to recover a reward function from expert demonstration data (Arora & Doshi, 2021; Ng et al., 2000). With advances in pretrained foundation models, some recent works have also studied using large language models or vision-language models to provide reward signals (Ma et al., 2022; Fan et al., 2022; Du et al., 2023; Karamcheti et al., 2023; Kwon et al., 2023; Wang et al., 2024; Ma et al., 2024). Reward Design Among these approaches, EUREKA (Ma et al., 2023) is the closest to our work, instructing the LLM to generate and select novel reward functions based on environment feedback with an evolutionary framework. However, as opposed to performing preference learning, EUREKA uses the LLM to design dense rewards that help with learning a hand-designed sparse reward. In contrast, ICPL focuses on finding a reward that correctly orders a set of preferences. However, we note that EUREKA also has a small, preliminary investigation combining human preferences with an LLM to generate human-preferred behaviors in a single scenario. We note that this experiment had the human provide their feedback in text, whereas we only use preference queries. This paper is a significantly scaled-up version of that investigation as well as a methodological study of how best to incorporate prior rounds of feedback. Human-in-the-loop Reinforcement Learning. Feedback from humans has been proven to be effective in training reinforcement learning agents that better match human preferences (Retzlaff et al., 2024; Mosqueira-Rey et al., 2023; Kwon et al., 2023). Previous works have investigated human feedback in various forms, such as trajectory comparisons, preferences, demonstrations, and corrections (Wirth et al., 2017; Ng et al., 2000; Jeon et al., 2020; Peng et al., 2024). Among these various methods, preference-based RL has been successfully scaled to train large foundation models for hard tasks like dialogue, e.g. ChatGPT (Ouyang et al., 2022). In LLM-based applications, prompting is a simple way to provide human feedback in order to align LLMs with human preferences (Giray, 2023; White et al., 2023; Chen et al., 2023). Iteratively refining the prompts with feedback from the environment or human users has shown promise in improving the output of the LLM (Wu et al., 2021; Nasiriany et al., 2024). This work extends the usage of the ability to control LLM behavior via in-context prompts. We aim to utilize interactive rounds of preference feedback between the LLM and humans to guide the LLM to generate reward functions that can elicit behaviors that align with human preferences. Human-in-the-loop Reinforcement Learning. Authors: (1) Chao Yu, Tsinghua University; (2) Hong Lu, Tsinghua University; (3) Jiaxuan Gao, Tsinghua University; (4) Qixin Tan, Tsinghua University; (5) Xinting Yang, Tsinghua University; (6) Yu Wang, with equal advising from Tsinghua University; (7) Yi Wu, with equal advising from Tsinghua University and the Shanghai Qi Zhi Institute; (8) Eugene Vinitsky, with equal advising from New York University (zoeyuchao@gmail.com). Authors: Authors: (1) Chao Yu, Tsinghua University; (2) Hong Lu, Tsinghua University; (3) Jiaxuan Gao, Tsinghua University; (4) Qixin Tan, Tsinghua University; (5) Xinting Yang, Tsinghua University; (6) Yu Wang, with equal advising from Tsinghua University; (7) Yi Wu, with equal advising from Tsinghua University and the Shanghai Qi Zhi Institute; (8) Eugene Vinitsky, with equal advising from New York University ( zoeyuchao@gmail.com ). zoeyuchao@gmail.com This paper is available on arxiv under CC 4.0 license. This paper is available on arxiv under CC 4.0 license. available on arxiv

Hacking Reinforcement Learning with a Little Help from Humans (and LLMs)

Could Science Benefit From AI-mediated Communication? This Study Says it Could

Author:
(1) David M. Markowitz, Department of Communication, Michigan State University, East Lansing, MI 48824. Editor's note: This is part 8 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Table of Links Abstract
The Benefits of Simple Writing
The Current Work
Study 1a: Method
Study 1a: Results
Study 1b: Method
Study 1b: Results
Study 2: Method
Study 2: Results
General Discussion, Acknowledgements, and References Study 2: Method Participants in the US were recruited from Prolific and paid $4.00 for their time in a short study (median completion time < 7 minutes). People were told that they would read scientific summaries and make judgments about the authors of such texts. Participants and Power Based on this study’s preregistration (https://aspredicted.org/C3K_T31), a minimum of 164 participants were required to detect a small effect powered at 80% in a within-subjects study (f = 0.10, α = two-tailed, three measurements). A total of 274 participants were recruited to ensure enough participants were in the study. Most participants self-identified as men (n = 139; 50.7%; women n = 127, other n = 7), they were 36.74 years old on average (SD = 12.47 years), and were mostly White (n = 190; 69.3%). On a 7-point political ideology scale (1 = extremely liberal, 7 = extremely conservative), participants leaned liberal (M = 2.97, SD = 1.63). Procedure Five pairs of stimuli from Study 1b were selected for the experiment, having had the greatest difference in common words scores between the PNAS and GPT texts (Pair 1 GPT = 79.31%, Pair 1 PNAS = 48.65%; Pair 2 GPT = 76.64%, Pair 2 PNAS = 46.32%; Pair 3 GPT = 79.07%, Pair 3 PNAS = 52.14%; Pair 4 GPT = 85.00%, Pair 4 PNAS = 59.66%; Pair 5 GPT = 87.10%, Pair 5 PNAS = 62.81%). Participants were randomly assigned to read stimuli from three out of a possible five pairs (see the online supplement for the stimuli texts), and within these randomly selected pairs, participants were randomly assigned to the GPT (simple) or PNAS (complex) version of each pair. Participants were told to read each summary of a scientific paper and then answer questions below each summary. They were specifically told “we are not expecting you to be an expert in the topic discussed below. Instead, make your judgments based on how the summary is written.” Finally, participants made various perceptions of the author (e.g., intelligence, trustworthiness) based on prior work (14, 15, 34), judgments about the identity of who wrote the scientific summary (AI or human), and assessed the complexity in each text as a manipulation check. The order of these measures was randomized, and items within each block were randomized as well. This study was approved by the author’s university research ethics board. Measures Manipulation Check Based on prior work (15, 34), three questions asked participants to rate how clear (“How clear was the writing in the summary you just read?”), complex (“How complex was the writing in the summary you just read?”), and how well they understood each scientific summary (“How much of this writing did you understand?”). Ratings for the first two questions were made on 7- point Likert-type scales from 1 = Not at all to 7 = Extremely. The third question ranged from 1 = Not at all to 7 = An enormous amount. Author Perceptions Participants made three ratings about the author of each scientific summary: (1) “How intelligent is the scientist who wrote this summary?”, (2) How credible is the scientist who wrote this summary?”, and (3) “How trustworthy is the scientist who wrote this summary?” As a collection, these dimensions were highly reliable (Cronbach’s α = 0.88) and therefore, they were averaged to create a general author perceptions index, while also being evaluated individually. All items were measured on 7-point Likert-type scales from 1 = Not at all to 7 = Extremely. Author Identity Perceptions Participants were asked for their agreement with two questions: (1) This summary was written by a human, and (2) This summary was written by Artificial Intelligence. All items were measured on 7-point Likert-type scales from 1 = Strongly disagree to 7 = Strongly agree. Demographics Basic demographic data were obtained from each participant, including their age, gender, ethnicity, and political ideology. Analytic Plan Since there were multiple observations per participant, linear mixed models with random intercepts for participant and stimulus were constructed (45, 46). This paper is available on arxiv under CC BY 4.0 DEED license. Author: (1) David M. Markowitz, Department of Communication, Michigan State University, East Lansing, MI 48824. Author: Author: (1) David M. Markowitz, Department of Communication, Michigan State University, East Lansing, MI 48824. Editor's note: This is part 8 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Editor's note: This is part 8 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Editor's note: This is part 8 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Table of Links Abstract The Benefits of Simple Writing The Current Work Study 1a: Method Study 1a: Results Study 1b: Method Study 1b: Results Study 2: Method Study 2: Results General Discussion, Acknowledgements, and References Abstract Abstract The Benefits of Simple Writing The Benefits of Simple Writing The Current Work The Current Work Study 1a: Method Study 1a: Method Study 1a: Results Study 1a: Results Study 1b: Method Study 1b: Method Study 1b: Results Study 1b: Results Study 2: Method Study 2: Method Study 2: Results Study 2: Results General Discussion, Acknowledgements, and References General Discussion, Acknowledgements, and References Study 2: Method Participants in the US were recruited from Prolific and paid $4.00 for their time in a short study (median completion time < 7 minutes). People were told that they would read scientific summaries and make judgments about the authors of such texts. Participants and Power Participants and Power Based on this study’s preregistration (https://aspredicted.org/C3K_T31), a minimum of 164 participants were required to detect a small effect powered at 80% in a within-subjects study (f = 0.10, α = two-tailed, three measurements). A total of 274 participants were recruited to ensure enough participants were in the study. Most participants self-identified as men (n = 139; 50.7%; women n = 127, other n = 7), they were 36.74 years old on average (SD = 12.47 years), and were mostly White (n = 190; 69.3%). On a 7-point political ideology scale (1 = extremely liberal, 7 = extremely conservative), participants leaned liberal (M = 2.97, SD = 1.63). Procedure Procedure Five pairs of stimuli from Study 1b were selected for the experiment, having had the greatest difference in common words scores between the PNAS and GPT texts (Pair 1 GPT = 79.31%, Pair 1 PNAS = 48.65%; Pair 2 GPT = 76.64%, Pair 2 PNAS = 46.32%; Pair 3 GPT = 79.07%, Pair 3 PNAS = 52.14%; Pair 4 GPT = 85.00%, Pair 4 PNAS = 59.66%; Pair 5 GPT = 87.10%, Pair 5 PNAS = 62.81%). Participants were randomly assigned to read stimuli from three out of a possible five pairs (see the online supplement for the stimuli texts), and within these randomly selected pairs, participants were randomly assigned to the GPT (simple) or PNAS (complex) version of each pair. Participants were told to read each summary of a scientific paper and then answer questions below each summary. They were specifically told “we are not expecting you to be an expert in the topic discussed below. Instead, make your judgments based on how the summary is written.” Finally, participants made various perceptions of the author (e.g., intelligence, trustworthiness) based on prior work (14, 15, 34), judgments about the identity of who wrote the scientific summary (AI or human), and assessed the complexity in each text as a manipulation check. The order of these measures was randomized, and items within each block were randomized as well. This study was approved by the author’s university research ethics board. Measures Measures Manipulation Check Manipulation Check Manipulation Check Based on prior work (15, 34), three questions asked participants to rate how clear (“How clear was the writing in the summary you just read?”), complex (“How complex was the writing in the summary you just read?”), and how well they understood each scientific summary (“How much of this writing did you understand?”). Ratings for the first two questions were made on 7- point Likert-type scales from 1 = Not at all to 7 = Extremely. The third question ranged from 1 = Not at all to 7 = An enormous amount. Author Perceptions Author Perceptions Author Perceptions Participants made three ratings about the author of each scientific summary: (1) “How intelligent is the scientist who wrote this summary?”, (2) How credible is the scientist who wrote this summary?”, and (3) “How trustworthy is the scientist who wrote this summary?” As a collection, these dimensions were highly reliable (Cronbach’s α = 0.88) and therefore, they were averaged to create a general author perceptions index, while also being evaluated individually. All items were measured on 7-point Likert-type scales from 1 = Not at all to 7 = Extremely. Author Identity Perceptions Author Identity Perceptions Author Identity Perceptions Participants were asked for their agreement with two questions: (1) This summary was written by a human, and (2) This summary was written by Artificial Intelligence. All items were measured on 7-point Likert-type scales from 1 = Strongly disagree to 7 = Strongly agree. Demographics Demographics Demographics Basic demographic data were obtained from each participant, including their age, gender, ethnicity, and political ideology. Analytic Plan Analytic Plan Since there were multiple observations per participant, linear mixed models with random intercepts for participant and stimulus were constructed (45, 46). This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Are AI or Humans More Trustworthy? A Study Puts Readers to the Test

Author:
(1) David M. Markowitz, Department of Communication, Michigan State University, East Lansing, MI 48824. Editor's note: This is part 7 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Table of Links Abstract
The Benefits of Simple Writing
The Current Work
Study 1a: Method
Study 1a: Results
Study 1b: Method
Study 1b: Results
Study 2: Method
Study 2: Results
General Discussion, Acknowledgements, and References Study 1b: Results Distributions of the comparisons in this study are reflected in Figure 1. Indeed, GPT significance statements were written in a simpler manner than PNAS significance statements for the simplicity index, Welch’s t(1492.1) = 11.55, p < .001, Cohen’s d = 0.58, 95% CI [0.47, 0.69]. Specifically, GPT significance statements (M = 75.53%, SD = 5.57%) contained more common words than PNAS significance statements (M = 69.84%, SD = 7.45%), Welch’s t(1478.7) = 17.31, p < .001, Cohen’s d = 0.87, 95% CI [0.76, 0.97]. GPT significance statements (M = 17.59, SD = 11.15) were also more readable than PNAS significance statements (M = 12.86, SD = 14.27), Welch’s t(1510) = 7.39, p < .001, Cohen’s d = 0.37, 95% CI [0.27, 0.47]. However, GPT significance statements (M = 92.73, SD = 6.89) had a statistically equivalent analytic style as PNAS significance statements (M = 92.32, SD = 7.48), Welch’s t(1587.7) = 1.16, p = .246, Cohen’s d = 0.06, 95% CI [-0.04, 0.16]. All results were maintained when comparing GPT significance statements to PNAS abstracts and PNAS significance statements as well. Alternative Explanations One possible explanation for the Study 1b results is that there are content differences across the PNAS and GPT texts explaining or impacting such differences across groups. This concern was addressed in two ways. First, PNAS has various sections that authors submit to, and LIWC has categories to approximate words associated with such sections. For example, the LIWC category for political speech would approximate papers submitted the Social Science section, specifically Political Sciences. Several linguistic covariates were therefore examined to account for content-related differences across GPT and PNAS texts. After including overall affect/emotion and cognition (to control for topics within the Psychological Sciences section of PNAS), political speech (to control for topics within the Political Science section of PNAS), and physical references to the multivariate models (to control for topics within the Biological Sciences section of PNAS), all results were maintained except for Analytic writing, where GPT texts were more analytic than PNAS texts, which is also consistent with prior work (42). Please see the online supplement for additional LIWC differences across these text types. Content effects were also evaluated in a bottom-up manner using the Meaning Extraction Method to measure dominant themes across the GPT and PNAS texts (43, 44). The evidence in the online supplement states there were 8 themes reliably extracted from the data, ranging from basic methodological and research information to gene expression and cancer science. Controlling for these themes, including the prior LIWC content dimensions, revealed consistent results as well (see supplement). Therefore, Study 1b evidence is robust to content. Altogether, human authors write simpler for lay audiences than for scientific audiences (Study 1a), but Study 1b demonstrated artificial intelligence and large language models can do so more effectively (e.g., the effect size differences between GPT significance statements and PNAS significance statements was larger than human in Study 1a). The findings thus far are correlational and therefore need causal evidence to demonstrate the impact of these effects on human perceptions. In Study 2, participants were randomly assigned to read a GPT significance statement or PNAS significance statement from pairs of texts that appeared in the previous studies. Participants made perceptions about the author (e.g., intelligence, credibility, trustworthiness), judged the complexity of each text, and they rated how much they believed the author of each text was human or artificial intelligence. Only perceptions of the author were made because prior work suggests people generally report consistent ratings when asked about both scientists and their science in similar studies (9). This paper is available on arxiv under CC BY 4.0 DEED license. Author: (1) David M. Markowitz, Department of Communication, Michigan State University, East Lansing, MI 48824. Author: Author: (1) David M. Markowitz, Department of Communication, Michigan State University, East Lansing, MI 48824. Editor's note: This is part 7 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Editor's note: This is part 7 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Editor's note: This is part 7 of 10 of a paper evaluating the effectiveness of using generative AI to simplify science communication and enhance public trust in science. The rest of the paper can be accessed via the table of links below. Table of Links Abstract The Benefits of Simple Writing The Current Work Study 1a: Method Study 1a: Results Study 1b: Method Study 1b: Results Study 2: Method Study 2: Results General Discussion, Acknowledgements, and References Abstract Abstract The Benefits of Simple Writing The Benefits of Simple Writing The Current Work The Current Work Study 1a: Method Study 1a: Method Study 1a: Results Study 1a: Results Study 1b: Method Study 1b: Method Study 1b: Results Study 1b: Results Study 2: Method Study 2: Method Study 2: Results Study 2: Results General Discussion, Acknowledgements, and References General Discussion, Acknowledgements, and References Study 1b: Results Distributions of the comparisons in this study are reflected in Figure 1. Indeed, GPT significance statements were written in a simpler manner than PNAS significance statements for the simplicity index, Welch’s t(1492.1) = 11.55, p < .001, Cohen’s d = 0.58, 95% CI [0.47, 0.69]. Specifically, GPT significance statements (M = 75.53%, SD = 5.57%) contained more common words than PNAS significance statements (M = 69.84%, SD = 7.45%), Welch’s t(1478.7) = 17.31, p < .001, Cohen’s d = 0.87, 95% CI [0.76, 0.97]. GPT significance statements (M = 17.59, SD = 11.15) were also more readable than PNAS significance statements (M = 12.86, SD = 14.27), Welch’s t(1510) = 7.39, p < .001, Cohen’s d = 0.37, 95% CI [0.27, 0.47]. However, GPT significance statements (M = 92.73, SD = 6.89) had a statistically equivalent analytic style as PNAS significance statements (M = 92.32, SD = 7.48), Welch’s t(1587.7) = 1.16, p = .246, Cohen’s d = 0.06, 95% CI [-0.04, 0.16]. All results were maintained when comparing GPT significance statements to PNAS abstracts and PNAS significance statements as well. Alternative Explanations Alternative Explanations One possible explanation for the Study 1b results is that there are content differences across the PNAS and GPT texts explaining or impacting such differences across groups. This concern was addressed in two ways. First, PNAS has various sections that authors submit to, and LIWC has categories to approximate words associated with such sections. For example, the LIWC category for political speech would approximate papers submitted the Social Science section, specifically Political Sciences. Several linguistic covariates were therefore examined to account for content-related differences across GPT and PNAS texts. After including overall affect/emotion and cognition (to control for topics within the Psychological Sciences section of PNAS), political speech (to control for topics within the Political Science section of PNAS), and physical references to the multivariate models (to control for topics within the Biological Sciences section of PNAS), all results were maintained except for Analytic writing, where GPT texts were more analytic than PNAS texts, which is also consistent with prior work (42). Please see the online supplement for additional LIWC differences across these text types. Content effects were also evaluated in a bottom-up manner using the Meaning Extraction Method to measure dominant themes across the GPT and PNAS texts (43, 44). The evidence in the online supplement states there were 8 themes reliably extracted from the data, ranging from basic methodological and research information to gene expression and cancer science. Controlling for these themes, including the prior LIWC content dimensions, revealed consistent results as well (see supplement). Therefore, Study 1b evidence is robust to content. Altogether, human authors write simpler for lay audiences than for scientific audiences (Study 1a), but Study 1b demonstrated artificial intelligence and large language models can do so more effectively (e.g., the effect size differences between GPT significance statements and PNAS significance statements was larger than human in Study 1a). The findings thus far are correlational and therefore need causal evidence to demonstrate the impact of these effects on human perceptions. In Study 2, participants were randomly assigned to read a GPT significance statement or PNAS significance statement from pairs of texts that appeared in the previous studies. Participants made perceptions about the author (e.g., intelligence, credibility, trustworthiness), judged the complexity of each text, and they rated how much they believed the author of each text was human or artificial intelligence. Only perceptions of the author were made because prior work suggests people generally report consistent ratings when asked about both scientists and their science in similar studies (9). This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

AI Crushes It at Simplicity: GPT-4 Writes Science Summaries Better Than the Pros

Stories, News & Mentions about HUMANA

Humana’s Stories on HackerNoon

Humana's News & Mentions