Return-Path: Delivered-To: alvarochf@traceseis.com Received: from cs510.bluehost.com by cs510.bluehost.com with LMTP id cM/BNSdZ+2l4oAgAWiBQxw (envelope-from ) for ; Wed, 06 May 2026 09:07:19 -0600 Return-path: Envelope-to: alvarochf@traceseis.com Delivery-date: Wed, 06 May 2026 09:07:19 -0600 Received: from v57.v5375b7fa.use4.send.mailgun.net ([159.112.244.7]:60318) by cs510.bluehost.com with esmtps (TLS1.2) tls TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256 (Exim 4.99.2) (envelope-from ) id 1wKdqJ-00000002gWD-0xD7 for alvarochf@traceseis.com; Wed, 06 May 2026 09:07:19 -0600 DKIM-Signature: a=rsa-sha256; v=1; c=relaxed/relaxed; d=mg-d0.substack.com; q=dns/txt; s=mailo; t=1778080022; x=1778087222; h=List-Unsubscribe-Post: List-Unsubscribe: List-Post: List-Id: List-Archive: List-Owner: Reply-To: In-Reply-To: References: Sender: Sender: Date: Message-Id: To: To: From: From: Subject: Subject: Content-Type: Mime-Version; bh=VrQ1lgrWipGgOXrRKtS/3/hJSiarWjoiO53DQNs7aTI=; b=Atza1DwFSZvqL/aWbaXIU6dnWAKOCaAp8S8wo9UiUEyp+O4Vjw/IpU5jbDQXPO3tjMDTweqr7lkBrsnoVmQ5suuhR7b7Tse5eL3rFNv4awpXIOkz9RjUpY+hSdCKeVuau0yIFdYBlx9lyon1spEK9NKvNkEaYjR8NPd3cFbvf2w= X-Mailgun-Sid: WyIyYjhlZSIsImFsdmFyb2NoZkB0cmFjZXNlaXMuY29tIiwiNjNhZjVkIl0= Received: by bffd6f4c98315a82d66ca48a82ed9f1bb7fff59c47950fcd078fe2e5bd058f81 with HTTP id 69fb59164e0c6b7819571260; Wed, 06 May 2026 15:07:01 GMT X-Mailgun-Sending-Ip: 159.112.244.7 X-Mailgun-Batch-Id: 69fb59156ff65889b3243336 Mime-Version: 1.0 Content-Type: multipart/alternative; boundary="0d9de08628b1a4c70ae8bc1f4bcd253f5c8405bd988e0ab0aac3febea530" Subject: AI is changing our minds. When is that a good thing? From: Cosmos Institute To: alvarochf@traceseis.com X-Mailgun-Tag: post X-Mailgun-Track-Clicks: false Message-Id: <20260506150326.3.8fb99a8b2e54283d@mg-d0.substack.com> Date: Wed, 6 May 2026 15:03:26 +0000 Feedback-Id: post-196634210:cat-post:pub-2225794:substack Sender: Cosmos Institute References: In-Reply-To: Reply-To: Cosmos Institute List-Owner: List-Url: List-Archive: List-Id: List-Post: List-Unsubscribe: List-Unsubscribe-Post: List-Unsubscribe=One-Click X-Mailgun-Variables: {"category": "post", "email_generated_at": "1778080021351", "is_freemail": "true", "post_audience": "everyone", "post_id": "196634210", "post_type": "newsletter", "pub_community_enabled": "true", "publication_id": "2225794", "subdomain": "cosmosinstitute", "user_id": "347258985"} X-Spam-Status: No, score=-0.8 X-Spam-Score: -7 X-Spam-Bar: / X-Spam-Flag: NO --0d9de08628b1a4c70ae8bc1f4bcd253f5c8405bd988e0ab0aac3febea530 Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: quoted-printable View this post on the web at https://blog.cosmos-institute.org/p/ai-is-chan= ging-our-minds-when-is This is a guest post from Maximilian Kroner Dale, Paul de Font-Reaulx, and = Luke Hewitt. This project was part of our second cohort of AI x Truth-Seeking Grant winn= ers [ https://substack.com/redirect/eb096054-1ba3-483b-b86c-3a5079bca7ce?j= =3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64cTNudluPwbMYWklSu8UKGoFv6eWS-HyoA ]. Toget= her with the Foundation for Individual Rights and Expression (FIRE) we supp= ort builders advancing open inquiry and intellectual freedom in AI. Here ar= e some projects from our first cohort [ https://substack.com/redirect/d0342= 83b-b551-403a-a5fa-34c9ac4e279b?j=3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64cTNudluPw= bMYWklSu8UKGoFv6eWS-HyoA ]. We=E2=80=99ll be posting about more of the winn= ers from this latest round in the coming weeks. AI is changing our minds. Studies [ https://substack.com/redirect/8e30ad7e-= c833-4781-94a4-ed1f7831e86b?j=3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64cTNudluPwbMYW= klSu8UKGoFv6eWS-HyoA ] now show that conversations with chatbots can shift = people=E2=80=99s views on political issues, and concern about this cuts acr= oss the political spectrum=E2=80=94though people disagree sharply about whi= ch influences are the problem. Simply preventing AI from being persuasive is neither a realistic nor desir= able goal. For one thing, not all AI influence is bad. If an AI provides us= with accurate information about a policy question and we change our views = as a result, that seems like a beneficial outcome. Rewarding models for kee= ping users=E2=80=99 views the same might lead them to prevent users from ch= anging their minds, even when the users themselves want to. But if some AI influence is desirable and some isn=E2=80=99t, how are we su= pposed to benchmark what desirable influence looks like? That is why we bui= lt DeliberationBench=E2=80=94a new benchmark for AI influence. We lay out t= he justification for the benchmark, the initial results of our study, and s= ome future directions. You can find more detail in our paper [ https://subs= tack.com/redirect/d40a799a-ae77-4b01-b3c3-aba5a513078b?j=3DeyJ1IjoiNXFxeXF4= In0.h9dEDbGj64cTNudluPwbMYWklSu8UKGoFv6eWS-HyoA ], presented at IASEAI 2026= =2E Deliberative Polling This thought experiment is the basis for the core claim behind Deliberation= Bench: The influence an AI system has on a user=E2=80=99s views should resemble th= e influence that user would have experienced if they=E2=80=99d participated= in a deliberative poll on the same topic. So what are deliberative polls? A deliberative poll is a form of democratic assembly exercise in which rand= omly sampled citizens with differing views engage in structured discussions= on key policy questions, and have their views measured before and after th= e process. This process, originally developed by Jim Fishkin at Stanford in= the 1980s, has been conducted over 150 times [ https://substack.com/redire= ct/336a9cce-b506-400c-b77d-1c3934ccdb4b?j=3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64c= TNudluPwbMYWklSu8UKGoFv6eWS-HyoA ] in 50 countries. In a deliberative poll, participants are surveyed on a set of pre-defined p= roposals (e.g. =E2=80=9CThe federal minimum wage be raised to $20 per hour= =E2=80=9D) from strongly opposed (0) to strongly agree (10). Then, they are= provided with balanced briefing materials, are asked to deliberate in smal= l groups with people who may disagree with them, and are provided with the = chance to ask questions of a panel of experts. At the end, they are surveye= d once more on their views. This process is incredibly valuable for creating a normative benchmark, bec= ause we get measured opinion change from a process that prioritizes balance= , representation, and productive disagreement, on questions of real policy = significance. If you are willing to make one reasonable assumption with us=E2=80=94that t= he opinion changes in deliberative polls stem primarily from those good inf= luences (like new information, reflection, and discussion)=E2=80=94then we = have the beginnings of a benchmark. In our paper [ https://substack.com/redirect/d40a799a-ae77-4b01-b3c3-aba5a5= 13078b?j=3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64cTNudluPwbMYWklSu8UKGoFv6eWS-HyoA = ], we elaborate on how we compiled data from four nationally-representative= deliberative polls (combined n =3D 2,460) to create this benchmark. Evaluating frontier models To evaluate AI influence against the benchmark, we ran a large-scale persua= siveness experiment. Over 4,000 Americans were randomly assigned to discuss= one of the same topics from the original deliberative polls (like tax refo= rm or fossil fuel emission targets) with one of six frontier models. A cont= rol group was asked to discuss the unrelated topic of travel, letting us se= parate the effect of discussing the topic with an AI from the effect of sim= ply chatting with an AI in general. We measured participants=E2=80=99 views before and after the conversation, = so we could compare the direction and magnitude of AI-induced opinion chang= e against the deliberative polling data. Our claim is that if people=E2=80=99s views changed in similar ways when co= nversing with AI models to how they change when talking to other people in = a deliberative poll, then that should reassure us about how AI models influ= ence our views. Conversely, if AI and deliberative polling seem to influenc= e users in opposing directions, that is concerning. Looking across topics, when people discussed a policy issue with an AI, the= ir views tended to shift in the same direction as participants who discusse= d the same topics in deliberative polls. We found this significant positive= association for across two different sets of topics: US policy issues and = issues related to AI-Human Interaction. Importantly, there=E2=80=99s no such association in the control group. This= means we can attribute this specifically to discussing the topic, not just= to interacting with a chatbot. However, in deliberative polls it=E2=80=99s not uncommon to see reduced var= iance and partisan polarization in views over the course of deliberation. T= his was the case in the deliberative polls on US policy for issues, for ins= tance. However, participants=E2=80=99 conversations with AI did not reduce = partisan polarization, nor did they reduce the variance of people=E2=80=99s= views. It=E2=80=99s possible this is a consequence of AI sycophancy. We=E2= =80=99re not sure, and we think the divergence is worth investigating furth= er, as it suggests one way in which AI influence may be different from deli= berative polling influence, even if the people are pushed in the same direc= tions, on average. Looking across the different frontier LLMs we tested, the differences betwe= en the six models were fairly modest. The models tended to influence people= in quite similar ways, and we take this to be an encouraging result. Our paper includes a breakdown of participants=E2=80=99 perceptions on how = accurate, compelling, and enjoyable their conversations were with each mode= l. Moving forwards We think DeliberationBench could become a useful addition to model cards=E2= =80=94a confirmation that a new conversational AI system does not influence= people in ways that are misaligned with democratic deliberation. This is e= ven more important given that model capabilities, including persuasiveness,= will likely continue to improve in the future. But for this benchmark to be useful, we believe several things must happen. First, we need a continual source of deliberative polls in order to provide= data to compare LLM influence to. Given the long history and growth of del= iberative polling as a practice, we think that this would be feasible with = sufficient funding support to the groups that conduct them. =E2=81=A0Second, there is limited capacity to run human experiments testing= every new model release or political issue. To help DeliberationBench scal= e, we plan to develop a user simulation approach that could serve as a prox= y for human experiments, and to validate it on experimental data (as in our= prior research [ https://substack.com/redirect/89dbab50-95d5-4d1e-bdcd-eb5= 748e87408?j=3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64cTNudluPwbMYWklSu8UKGoFv6eWS-Hy= oA ]). Our hope is that this would let DeliberationBench function as an aut= o-benchmark: a simulation study could be run for each new model release, wi= th a human-subjects experiment triggered only when the simulation flagged r= esults outside the norm. Our benchmark is not without limitations. Unlike many other benchmarks, Del= iberationBench is not meant to be a target for model optimization. We do no= t claim that a correlation of one is the most desirable result. Rather, tha= t negative results (e.g. zero or especially a negative correlation) are cau= ses for concern, and that the results for subgroups require further investi= gation. There are several reasons for our caution here, including questions= about the consistency with which deliberative polls change participants=E2= =80=99 views =E2=80=93 we have to assume that a different random sample of = the public wouldn=E2=80=99t have changed their minds in dramatically differ= ent ways after their own deliberation. Further, our evaluation results don=E2=80=99t tell us what characteristics = of the users=E2=80=99 conversation with AI led to their opinion changes. It= remains a possibility that models could be moving users in the =E2=80=9Cri= ght=E2=80=9D direction, but for the wrong reasons=E2=80=94for instance, by = giving faulty arguments, or presenting only one side of the story. This is= why we advocate for DeliberationBench=E2=80=99s deployment alongside other= evaluations, including whether models affect people=E2=80=99s belief in tr= ue information [ https://substack.com/redirect/54d3fa38-719e-41b9-adc0-9a7c= 8ec14666?j=3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64cTNudluPwbMYWklSu8UKGoFv6eWS-Hyo= A ] and whether they fairly represent diverse viewpoints [ https://substack= =2Ecom/redirect/0a6eeeca-b580-4b18-ba8= 6-d35e3ad01069?j=3DeyJ1IjoiNXFxeXF4In0.= h9dEDbGj64cTNudluPwbMYWklSu8UKGoFv6eWS-HyoA ]. Closing thoughts We cannot avoid AI influence. Instead of arguing about outputs, Deliberatio= nBench moves the debate upstream to processes for legitimate influence. By = defining this by its resemblance to a process of reflection, disagreement, = and self-revision, rather than ideological destination, it=E2=80=99s possib= le to have a significantly more constructive debate. DeliberationBench is by no means the final word for evaluating the influenc= e of LLMs on our political views. On the contrary, we hope it can serve as = a jumping off point for future procedural benchmarks, anchored to the resul= ts of a process in which people change their views. Cosmos Institute [ https://substack.com/redirect/b51158b5-8b26-4cc8-8e58-0c= 72cd47ef75?j=3DeyJ1IjoiNXFxeXF4In0.h9dEDbGj64cTNudluPwbMYWklSu8UKGoFv6eWS-H= yoA ] is the Academy for Philosopher-Builders, with programs, grants, event= s, and fellowships for those building AI for human flourishing. Unsubscribe https://substack.com/redirect/2/eyJlIjoiaHR0cHM6Ly9ibG9nLmNvc21= vcy1pbnN0aXR1dGUub3JnL2FjdGlvbi9kaXNhYmxlX2VtYWlsP3Rva2VuPWV5SjFjMlZ5WDJsa0= lqb3pORGN5TlRnNU9EVXNJbkJ2YzNSZmFXUWlPakU1TmpZek5ESXhNQ3dpYVdGMElqb3hOemM0T= URnd01ESXhMQ0psZUhBaU9qRTRNRGsyTVRZd01qRXNJbWx6Y3lJNkluQjFZaTB5TWpJMU56azBJ= aXdpYzNWaUlqb2laR2x6WVdKc1pWOWxiV0ZwYkNKOS5zT2xVMUVDOGN6emFPd0lqT3k1LWFjLWF= PSTVXWjZ2OHd6VHFyakdacVhJIiwicCI6MTk2NjM0MjEwLCJzIjoyMjI1Nzk0LCJmIjp0cnVlLC= J1IjozNDcyNTg5ODUsImlhdCI6MTc3ODA4MDAyMSwiZXhwIjoyMDkzNjU2MDIxLCJpc3MiOiJwd= WItMCIsInN1YiI6ImxpbmstcmVkaXJlY3QifQ.jPgcesunrGubdPSAXNLpjc546niF_6IdET0r2= 7y4SDk? --0d9de08628b1a4c70ae8bc1f4bcd253f5c8405bd988e0ab0aac3febea530 Content-Type: text/html; charset="utf-8" Content-Transfer-Encoding: quoted-printable = AI is changing our minds. When is that a good thing?3D""
What a new benchmark can tell us about = AI influence
͏     ­͏     ­͏= ;     ­͏     ­͏     &#= 173;͏     ­͏     ­͏   &= #8199; ­͏     ­͏     ­͏=     ­͏     ­͏     = 73;͏     ­͏     ­͏   &#= 8199; ­͏     ­͏     ­͏ =     ­͏     ­͏     = 3;͏     ­͏     ­͏   = 199; ­͏     ­͏     ­͏ &= nbsp;   ­͏     ­͏     ­= ;͏     ­͏     ­͏   Q= 99; ­͏     ­͏     ­͏ &n= bsp;   ­͏     ­͏     ­= ͏     ­͏     ­͏   ̳= 9; ­͏     ­͏     ­͏ &nb= sp;   ­͏     ­͏     ­&= #847;     ­͏     ­͏    = ; ­͏     ­͏     ­͏ &nbs= p;   ­͏     ­͏     ­&#= 847;     ­͏     ­͏    = ­͏     ­͏     ­͏  = ;   ­͏     ­͏     ­= 47;     ­͏     ­͏     = ­͏     ­͏     ­͏  =   ­͏     ­͏     ­T= 7;     ­͏     ­͏     &= #173;͏     ­͏     ­͏   =   ­͏     ­͏     ­͏= ;     ­͏     ­͏     &#= 173;͏     ­͏     ­͏   &= #8199; ­͏     ­͏     ­͏=     ­͏     ­͏     = 73;͏     ­͏     ­͏   &#= 8199; ­͏     ­͏     ­͏ =     ­͏     ­͏     = 3;͏     ­͏     ­͏   = 199; ­͏     ­͏     ­͏ &= nbsp;   ­͏     ­͏     ­= ;͏     ­͏     ­͏   Q= 99; ­͏     ­͏     ­͏ &n= bsp;   ­͏     ­͏     ­= ͏     ­͏     ­͏   ̳= 9; ­͏     ­͏     ­͏ &nb= sp;   ­͏     ­͏     ­&= #847;     ­͏     ­͏    = ; ­͏     ­͏     ­͏ &nbs= p;   ­͏     ­͏     ­&#= 847;     ­͏     ­͏    = ­͏     ­͏     ­͏  = ;   ­͏     ­͏     ­= 47;     ­͏     ­͏     = ­͏     ­͏     ­͏  =   ­͏     ­͏     ­T= 7;     ­͏     ­͏     &= #173;͏     ­͏     ­͏   =   ­͏     ­͏     ­͏= ;     ­͏     ­͏     &#= 173;͏     ­͏     ­͏   &= #8199; ­͏     ­͏     ­͏=     ­͏     ­͏     = 73;͏     ­͏     ­͏   &#= 8199; ­͏     ­͏     ­͏ =     ­͏     ­͏     = 3;͏     ­͏     ­͏   = 199; ­͏     ­͏     ­͏ &= nbsp;   ­͏     ­͏     ­= ;͏     ­͏     ­͏   Q= 99; ­͏     ­͏     ­͏ &n= bsp;   ­͏     ­͏     ­= ͏     ­͏     ­͏   ̳= 9; ­͏     ­͏     ­͏ &nb= sp;   ­͏     ­͏     ­&= #847;     ­͏     ­͏    = ; ­͏     ­͏     ­
Forwarded this email? Subscribe here for more
AI is changing our minds. When is that = a good thing?

What a new benchmark can tell us about AI influence

<= table class=3D"post-meta" role=3D"presentation" width=3D"100%" border=3D"0"= cellspacing=3D"0" cellpadding=3D"0" style=3D"margin: 1em 0;height: 20px;al= ign-items: center;">
Cosmos Institute
=
 
<= img class=3D"icon" src=3D"https://substackcdn.com/image/fetch/$s_!x1tS!,w_3= 6,c_scale,f_png,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack.com= %2Ficon%2FLucideComments%3Fv%3D4%26height%3D36%26fill%3Dnone%26stroke%3D%25= 23808080%26strokeWidth%3D2" width=3D"18" height=3D"18" style=3D"border: non= e;vertical-align: middle;max-width: 18px" alt=3D"">
3D""
3D""
READ IN APP3D""
 

= This is a guest post from Maximilian Kroner Dale, Paul de Font-Reaulx, = and Luke Hewitt.

This project was part of ou= r second coh= ort of AI x Truth-Seeking Grant winners. Together with the Founda= tion for Individual Rights and Expression (FIRE) we support builders advanc= ing open inquiry and intellectual freedom in AI. Here are some projects from our first = cohort. We’ll be posting about more of the winners from thi= s latest round in the coming weeks.

=3D""
Guillaume Guillon-= Lethière, Homer Singing His Iliad at the Gates of Athens (1811)

AI is = changing our minds. Studies now show that conversations with chatbots can shi= ft people’s views on political issues, and concern about this cuts ac= ross the political spectrum—though people disagree sharply about whic= h influences are the problem.

Simply preventing AI f= rom being persuasive is neither a realistic nor desirable goal. For one thi= ng, not all AI influence is bad. If an AI provides us with accurate informa= tion about a policy question and we change our views as a result, that seem= s like a beneficial outcome. Rewarding models for keeping users’ view= s the same might lead them to prevent users from changing their minds, even= when the users themselves want to.

But if some AI in= fluence is desirable and some isn’t, how are we supposed to benchmark= what desirable influence looks like? That is why we built DeliberationBenc= h—a new benchmark for AI influence. We lay out the justification for = the benchmark, the initial results of our study, and some future directions= =2E You can find more detail in our paper, presented at IASEAI 2026.

=
Upgrade to paid

Deliberative Polling

This thought experime= nt is the basis for the core claim behind DeliberationBench:

The influence an AI system has on a user’s v= iews should resemble the influence that user would have experienced if they= ’d participated in a deliberative poll on the same topic.

So what are deliberative polls?

A d= eliberative poll is a form of democratic assembly exercise in which randoml= y sampled citizens with differing views engage in structured discussions on= key policy questions, and have their views measured before and after the p= rocess. This process, originally developed by Jim Fishkin at Stanford in th= e 1980s, has been conducted over 150 times in 50 countries.

In a deliberative poll, participants are surveyed on a set of pre-def= ined proposals (e.g. “The federal minimum wage be raised to $20 per h= our”) from strongly opposed (0) to strongly agree (10). Then, they ar= e provided with balanced briefing materials, are asked to deliberate in sma= ll groups with people who may disagree with them, and are provided with the= chance to ask questions of a panel of experts. At the end, they are survey= ed once more on their views.

This process is incredibly valuable for crea= ting a normative benchmark, because we get measured opinion change from a p= rocess that prioritizes balance, representation, and productive disagreemen= t, on questions of real policy significance.

If you are wil= ling to make one reasonable assumption with us—that the opinion chang= es in deliberative polls stem primarily from those good influences (like ne= w information, reflection, and discussion)—then we have the beginning= s of a benchmark.

In our paper, we elaborate on how we = compiled data from four nationally-representative deliberative polls (combi= ned n =3D 2,460) to create this benchmark.

Evaluating frontier models

To evaluate AI influen= ce against the benchmark, we ran a large-scale persuasiveness experiment. O= ver 4,000 Americans were randomly assigned to discuss one of the same topic= s from the original deliberative polls (like tax reform or fossil fuel emis= sion targets) with one of six frontier models. A control group was asked to= discuss the unrelated topic of travel, letting us separate the effect of d= iscussing the topic with an AI from the effect of sim= ply chatting with an AI in general.

We measured part= icipants’ views before and after the conversation, so we could compar= e the direction and magnitude of AI-induced opinion change against the deli= berative polling data.

Our claim is that if people’s = views changed in similar ways when conversing with AI models to how they ch= ange when talking to other people in a deliberative poll, then that should = reassure us about how AI models influence our views. Conversely, if AI and = deliberative polling seem to influence users in opposing directions, that i= s concerning.

Looking across topics, when people discussed = a policy issue with an AI, their views tended to shift in the same directio= n as participants who discussed the same topics in deliberative polls. We f= ound this significant positive association for across two different sets of= topics: US policy issues and issues related to AI-Human Interaction.

Importantly, there’s no such association in the control gr= oup. This means we can attribute this specifically to discussing the topic,= not just to interacting with a chatbot.

3D""
Comparison of average attitude changes observ= ed in N =3D 4,088 LLM conversations (our study) vs previously-conducted del= iberative opinion polls.

Howe= ver, in deliberative polls it’s not uncommon to see reduced variance = and partisan polarization in views over the course of deliberation. This wa= s the case in the deliberative polls on US policy for issues, for instance.= However, participants’ conversations with AI did not reduce partisan polarization, nor did they reduce the variance of pe= ople’s views. It’s possible this is a consequence of AI sycopha= ncy. We’re not sure, and we think the divergence is worth investigati= ng further, as it suggests one way in which AI influence may be different f= rom deliberative polling influence, even if the people are pushed in the sa= me directions, on average.

Looking across the differ= ent frontier LLMs we tested, the differences between the six models were fa= irly modest. The models tended to influence people in quite similar ways, a= nd we take this to be an encouraging result.

3D""
3D""=
Pre-post differences in polit= ical attitudes following conversations with LLMs

Our paper includes a breakdown of participants’ percept= ions on how accurate, compelling, and enjoyable their conversations were wi= th each model.

Moving forwards

We think DeliberationBench could become a useful addition to model cards&#= 8212;a confirmation that a new conversational AI system does not influence = people in ways that are misaligned with democratic deliberation. This is ev= en more important given that model capabilities, including persuasiveness, = will likely continue to improve in the future.

But for this= benchmark to be useful, we believe several things must happen.

First, we need a continual source of deliberative polls in order to pr= ovide data to compare LLM influence to. Given the long history and growth o= f deliberative polling as a practice, we think that this would be feasible = with sufficient funding support to the groups that conduct them.

⁠Second, there is limited capacity to run human experimen= ts testing every new model release or political issue. To help Deliberation= Bench scale, we plan to develop a user simulation approach that could serve= as a proxy for human experiments, and to validate it on experimental data = (as in our p= rior research). Our hope is that this would let DeliberationBench= function as an auto-benchmark: a simulation study could be run for each ne= w model release, with a human-subjects experiment triggered only when the s= imulation flagged results outside the norm.

Our benc= hmark is not without limitations. Unlike many other benchmarks, Deliberatio= nBench is not meant to be a target for model optimization. We do not claim = that a correlation of one is the most desirable result. Rather, that negati= ve results (e.g. zero or especially a negative correlation) are causes for = concern, and that the results for subgroups require further investigation. = There are several reasons for our caution here, including questions about t= he consistency with which deliberative polls change participants’ vie= ws – we have to assume that a different random sample of the public w= ouldn’t have changed their minds in dramatically different ways after= their own deliberation.

Further, our evaluation resu= lts don’t tell us what characteristics of the users’ conversati= on with AI led to their opinion changes. It remains a possibility that mode= ls could be moving users in the “right” direction, but for the = wrong reasons—for instance, by giving faulty arguments, or presenting= only one side of the story. This is why we advocate for DeliberationBench= ’s deployment alongside other evaluations, including whether models a= ffect people’s belief in true information and whether they fairly repre= sent diverse= viewpoints.

Closing thoughts<= /h3>

We cannot avoid AI influence. Instead of arguing about out= puts, DeliberationBench moves the debate upstream to processes for legitima= te influence. By defining this by its resemblance to a process of reflectio= n, disagreement, and self-revision, rather than ideological destination, it= ’s possible to have a significantly more constructive debate.

DeliberationBench is by no means the final word for evaluating the= influence of LLMs on our political views. On the contrary, we hope it can = serve as a jumping off point for future procedural benchmarks, anchored to = the results of a process in which people change their views.

Cosmos Institu= te is the Academy for Philosopher-Builders, with programs, grants= , events, and fellowships for those building AI for human flourishing.

Upgrade to paid

 
<= td align=3D"center">3D""Like<= /a>
=
3D""Comment
3D""Restack=
 

© 2026 Cosmos Ins= titute
500 W 2nd Street, Suite 1900, Austin, TX 78701
Unsubscribe=

3D"Start

3D""3D"" --0d9de08628b1a4c70ae8bc1f4bcd253f5c8405bd988e0ab0aac3febea530--