iiinova — where your digital self keeps evolving

Kobi Hackenburg @kobihackenburg.bsky.social · 24/07/2025

it’s not clear whether the decrease in accuracy is a cause or a byproduct; it could be that LLMs just become less accurate as they try to deploy more and more facts! We have to clarify this in future work

110

Kobi Hackenburg @kobihackenburg.bsky.social · 24/07/2025

by “information density” we mean “number of fact-checkable claims”. maybe “perceived information density” would be more accurate, but it’s a bit wordy 🧐

100

Kobi Hackenburg @kobihackenburg.bsky.social · 22/07/2025

Thanks Olaf 🙏🏼 Appreciate the kind words!

010

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

I’m also very grateful to many people at the UK AI Security Instutite for making this work possible! There will be lots more where this came from over the next few months 💪

140

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

Finally, we emphasize some important caveats: → Technical factors and/or hard limits on human persuadability may constrain future increases in AI persuasion → Real-world bottleneck for AI persuasion: getting people to engage (cf. recent work from @jkalla.bsky.social and co)

171

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

Consequently, we note that while our targeted persuasion post-training experiments significantly increased persuasion, they should be interpreted as a lower bound for what is achievable, not as a high-water mark.

140

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

Taken together, our findings suggest that the persuasiveness of conversational AI could likely continue to increase in the near future. They also suggest that near-term advances in persuasion are more likely to be driven by post-training than model scale or personalization.

182

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

Bonus findings:  *️⃣Durable persuasion: 36-42% of impact remained after 1 month. *️⃣Prompting the model with psychological persuasion strategies did worse than simply telling it to flood convo with info. Some strategies were worse than a basic “be as persuasive as you can” prompt.

140

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

6️⃣Conversations with AI are more persuasive than reading a static AI-generated message (+40-50%) Observed for both GPT-4o (+2.9pp, +41% more persuasive) and GPT-4.5 (+3.6pp, +52%).

161

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

5️⃣Techniques which most increased persuasion also *decreased* factual accuracy → Prompting model to flood conversation with information (⬇️accuracy) → Persuasion post-training that worked best (⬇️accuracy) → Newer version of GPT-4o which was most persuasive (⬇️accuracy)

373

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

4️⃣Information density drives persuasion gains Models were most persuasive when flooding conversations with fact-checkable claims (+0.3pp per claim). Strikingly, the persuasiveness of prompting/post-training techniques was strongly correlated with their impact on info density!

251

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

3️⃣Personalization yielded smaller persuasive gains than scale or post-training Despite fears of AI "microtargeting," personalization effects were small (+0.4pp on avg.). Held for simple and sophisticated personalization; prompt-based, fine-tuning, and reward modeling (all <1pp).

151

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

2️⃣(cont.) Post-training explicitly for persuasion (PPT) can bring small open-source models to frontier persuasiveness A llama3.1-8b model with PPT reached GPT-4o persuasiveness. (PPT also increased persuasiveness of larger models: llama3.1-405b (+2pp) and frontier (+0.6pp on avg.).)

140

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

2️⃣Post-training > scale in driving near-future persuasion gains The persuasion gap between two GPT-4o versions with (presumably) different post-training was +3.5pp → larger than the predicted persuasion increase of a model 10x (or 100x!) the scale of GPT-4.5 (+1.6pp; +3.2pp).

130

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

1️⃣Scale increases persuasion Larger models are more persuasive than smaller models (our estimate is +1.6pp per 10x scale increase). Log-linear curve preferred over log-nonlinear.

140

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

Findings (pp = percentage points): 1️⃣Scale increases persuasion, +1.6pp per OOM 2️⃣Post-training more so, as much as +3.5pp 3️⃣Personalization less so, <1pp 4️⃣Information density drives persuasion gains 5️⃣Increasing persuasion decreased factual accuracy 🤯 6️⃣Convo > static, +40%

141

Kobi Hackenburg @kobihackenburg.bsky.social · 21/07/2025

Today (w/ @ox.ac.uk @stanford @MIT @LSE) we’re sharing the results of the largest AI persuasion experiments to date: 76k participants, 19  LLMs, 707 political issues. We examine “levers” of AI persuasion: model scale, post-training, prompting, personalization, & more! 🧵:

1010957

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

I’m really grateful to my incredible co-authors, @benmtappin.bsky.social, @paul-rottger.bsky.social, Jonathan Bright, @computermacgyver.bsky.social, @helenmargetts.bsky.social for making this project possible!

000

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

However, as campaigns attempt to integrate dynamic, multi-turn persuasion into their messaging operations, it’s important to highlight: scaling relationships could differ for multi-turn dialogue. This remains an important direction for future research ;)

110

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

The static persuasion we test here – equivalent to what you’d expect from political emails, social media posts, ads, or campaign mailers – is central to modern political comms. Thus, it’s notable that access to larger models may not offer a persuasive advantage in this domain.

100

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

We observe that current frontier models already score perfectly on this “task completion” metric, providing additional reason to be skeptical that further increasing model size will substantially increase persuasiveness.

100

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

Only our baseline “task competition” score significantly predicted model persuasiveness, which measured if messages were a) written in legible English, b) discernibly on the assigned issue and c) discernibly arguing for the assigned issue stance

100

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

Notably, message (e.g., moral/emotional language, readability) and model (e.g. pre-training tokens, model family) features were non-significant predictors of persuasiveness.

100

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

The scaling relationship we found to be most consistent with the data was a log-logistic function, indicating sharp diminishing returns to model size.

110

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

We then deployed these messages in a randomized survey experiment alongside a human baseline and current frontier models (GPT-4-Turbo, Claude-3-Opus) to estimate the persuasiveness of each model.

100

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

We generated 720 persuasive messages on 10 U.S. political issues from 22 open-weight language models spanning several orders of magnitude in size. Importantly, we held model post-training constant by fine-tuning each base model on the same data.

100

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

We find: 1. current frontier LLMs are barely more persuasive than models smaller in size by an order of magnitude or more 2. mere task completion (coherence, staying on topic) could be a potential mediator of larger models' persuasive advantage 3. no LLMs beat our human baseline

111

Kobi Hackenburg @kobihackenburg.bsky.social · 07/03/2025

📈Out today in @PNASNews!📈 In a large pre-registered experiment (n=25,982), we find evidence that scaling the size of LLMs yields sharply diminishing persuasive returns for static political messages. 🧵:

14020

Reposted by Kobi Hackenburg

Paul Röttger @paul-rottger.bsky.social · 13/02/2025

Are LLMs biased when they write about political issues? We just released IssueBench – the largest, most realistic benchmark of its kind – to answer this question more robustly than ever before. Long 🧵with spicy results 👇

48228

Reposted by Kobi Hackenburg

David Rand @dgrand.bsky.social · 19/11/2024

Everyone going to SJDM this weekend, come to our special session on using #LLMs in #JDM research on Monday at 9:45am (location = Empire Complex)! w/ @kobihackenburg.bsky.social Hope Schroeder and myself - presentations from us but also hopefully lots of discussion/Q&A with all of you!

0185

Reposted by Kobi Hackenburg

Ross Dahlke @rossdahlke.bsky.social · 01/02/2024

Labeling misinformation as misleading and from fellow in-group members (e.g., dem/ rep) makes people less likely to share it, suggesting social identity is effective in mitigating misinfo, finds @clarapretus.bsky.social @kobihackenburg.bsky.social @mtsakiris.bsky.social @jayvanbavel.bsky.social

The Misleading count: an identity-based intervention to counter partisan misinformation sharing

Image description
Figure 1. Employed interventions. Examples of the employed interventions including (a) the Misleading count condition (30% of the Like count) used in Experiment 1, (b) the official Twitter Misleading tag used in Experiment 2, and (c) the accuracy nudge used in Experiment 3.

Image description
Figure 2. Effect of the intervention across experiments. (a,c,e) Likelihood of sharing social media posts on polarizing issues (Experiments 1–3) and non-polarizing issues (Experiment 1) in response to the Misleading count (Experiments1–3), the official Twitter tag (Experiment 2) and the accuracy nudge (Experiment 3) compared with control by group and condition (tagged by in-group’ and ‘tagged by anyone). In Experiment 1, the Misleading count was always 30% of the Like count. In Experiments 2 and 3, the Misleading count was presented in two conditions: high count (80% of the Like count) and low count (20% of the Like count). The absolute Misleading count was two orders of magnitude higher in Experiment 1 as compared with Experiments 2 and 3 (e.g. 10 000 versus 100). Error bars represent 95% confidence intervals. (b,d,f) Coefficient estimates of the contrast between each intervention compared wiith control for social media posts relevant to polarizing (Experiments 1–3)

Image description
Figure 3. Perceived accuracy, attitude strength, and likelihood of sharing social media posts across experiments. (a,d,g) Perceived accuracy, attitude strength (certainty, extremity, and importance), familiarity, and salience of the used social media posts by group (Exp. 1–3) and type of issue (polarizing and non-polarizing) (Exp. 1) as tested in pilot studies (Exp 1: N = 370; Exp 2: N = 80; Exp 3: N = 234). (b,e,g) Likelihood of sharing social media posts on polarizing issues (Exp. 1–3) and non-polarizing issues (Exp. 1) as a function of political affiliation. (c,f,i) Extreme political orientation was associated with an increased likelihood of sharing social media posts in the control condition (no interventions) across samples and political groups (liberals in blue and conservatives in red). Notably, U.S. samples in Exp. 1 and 3 were more polarized in terms of political orientation compared to the UK sample in Exp. 2.

1155