Last weekend, the Financial Times published an article about the broader social impacts of AI. Studying, managing, and mitigating the negative impacts of the technological transition is supremely important, and doing so requires both the electorate and policymakers to be informed by good, impartial science. Unfortunately, for anyone with even basic training in social science research methods, the analysis by the FT’s John Burn-Murdoch was, at best, questionable.
His piece makes two claims. The first is that social media is inherently polarizing. According to the FT, these platforms “over-represent the radical right and left”, drive people to “conspiratorial beliefs” and political extremes, and are chiefly responsible for “waves of populism, polarization and an erosion of trust in experts, expertise and the establishment”.
The second, related claim is that AI chatbots — such as ChatGPT, Claude, and Grok — are the inverse, and will ultimately “nudge people away from the most extreme positions and towards more moderate and expert-aligned stances”. Burn-Murdoch labels ChatGPT, Gemini, and DeepSeek as “center-left”, and Grok as “center-right”. That he characterizes as “center-right” a model which famously referred to itself as “MechaHitler” and routinely peddles conspiracy theories about race and intelligence is probably worth interrogating elsewhere.
To justify the first claim, Burn-Murdoch references his own prior FT column. To justify the second, he conducted an experiment in which he created a series of “simulated users” (read: LLM instances), told half of them what they’re supposed to believe based on survey data and half not, and then had the simulated users discuss a range of political subjects with each of the mainstream LLMs. He then took the initial stated opinion of the simulated user, the response of the mainstream LLMs, and averaged them, applying an 80% weight to the initial position of the simulated user, and 20% to the LLM’s response.
He accounts for this approach by referencing a study jointly performed by the UK AI Security Institute, Oxford, LSE, Stanford, and MIT, published in the journal Science in December, on the subject of machine persuasion. Unfortunately, the study he cites does not even remotely use or justify any such methodology; instead, it tests the impact of human conversations with LLMs that are specifically tasked with persuading users. Nor do its results support an assumption of a universal 20% shift in user opinion in the direction of the model’s perceived base preferences. Rather, they evaluate multiple “persuasion strategies”, and then make an estimation about how far a model built for political persuasion could shift end user opinion.
This assumes, of course, that a user is open to being persuaded on contentious political subjects by a model, which can’t be generalized across the population to estimate aggregate societal impact. Moreover, where the cited article seeks to evaluate the impact of various persuasion strategies employed by LLMs on human users in an extended conversation, Burn-Murdoch seemingly just averages an LLM statement and simulated user statement together with an arbitrary weight, handwaving at “experimental evidence” as justification.
Additionally, the base claim that one can use an LLM trying to convince another LLM of an opinion as an accurate model of how an LLM could persuade a person remains entirely unjustified. Whether or not an LLM output indicates a genuine “change in belief” of that model or model persona is still an open question, particularly given models have a tendency to emphasize agreeableness and prioritize consensus. And while there is substantive interest in using simulated user behavior as a supplement or replacement for statistical surveys of real users, such methods are similarly unproven and methodologically tenuous.
But let us put even these concerns aside and return to the subject of convergent and divergent technologies of opinion. In the piece, Burn-Murdoch claims that LLMs will definitionally converge opinion, completely ignoring the well-trodden phenomenon of AI sycophancy, where models “excessively agree with, flatter, or validate” user beliefs. The fact that models broadly tell people what they want to hear, in ways that go far beyond social media “echo chambers”, in no way supports the claim that models will miraculously converge public opinion.
If we genuinely want to craft good AI policy, and have an informed debate on the role and impact of this technology on society, we cannot allow such shoddy work to pollute the commons, especially under the imprimatur of one of the most reputable and storied journalistic sources. Work like this risks marring that reputation when it comes to one of the most transformative issues of our era.