research
Research pack · 29 Oct 2026
Started from an Instagram exchange in which a creator who makes flower-decorated cookies (Loria Stern) said AI is free, and moved off the point when challenged. This is the frame for that moment, the evidence on what persuades, and the numbers behind the costs.
Dan Kahan, a professor at Yale Law School who leads the Cultural Cognition Project, describes a simple pattern: people tend to accept or dismiss evidence in ways that match the beliefs that dominate their group. The reason is not stupidity. For an individual, getting a big public question “wrong” costs almost nothing, while being out of step with friends, peers or customers can cost a lot. So the mind quietly protects belonging.
Kahan's 2017 synthesis argues that many public misconceptions depend on this dynamic, and that handing people correct information, on its own, rarely fixes it. The meaning attached to a fact (whose side it seems to be on) matters as much as the fact.
People judged whether someone was a credible expert partly by whether the expert's conclusion fit their own group's view. Across issues like climate change, guns and nuclear power, no cultural group was consistently better than another at recognising what scientists actually agreed on.
The people with the highest science literacy and numeracy were not the most worried about climate change. They were the most divided along cultural lines. Knowing more gave them more skill at defending their side.
Participants interpreted the same table of numbers. When it was about a skin cream, the most numerate people got it right most often. When the identical numbers were about a gun-control ban, answers split by politics, and the most numerate became the most polarised.
The same information about the vaccine moved people differently depending on whether the advocate seemed to share their values. Who said it mattered more than how emotionally it was said.
The same climate science was paired with different solutions. Readers told about geoengineering were less polarised about the evidence than readers told the answer was stricter emission limits, because the solution signalled whose values were under threat. Some critics question the study's design.
Unlike science knowledge, curiosity (enjoying being surprised by evidence) went with less polarisation. Curious people were much more willing to open a story that challenged their side.
Yale News on Kahan, Landrum, Carpenter, Helft and Jamieson, 2017
For a creator whose work is beauty, craft and nature, the online creative crowd tends to hold one of two group positions on AI: “it's a useful free tool” or “it's theft and waste”. A challenge like “AI isn't free” can land as “you're on the wrong side”, even if that wasn't the intent.
Arguments framed in the listener's own values move people more than arguments in the speaker's values. Studies in the US found, for example, that environmental messages framed around purity and patriotism moved conservatives more than framings about harm. Effects are real but usually modest.
Feinberg and Willer, “Moral reframing”, Social and Personality Psychology Compass, 2019 · Kahan et al., 2015
People trust information more when it comes from someone who seems to share their outlook. In the HPV experiment, the perceived values of the advocate shifted views more than the content did. A fellow baker, florist or food creator will be heard where an outsider is not.
Kahan et al., 2010 · Kahan, “Fixing the communications failure”, Nature, 2010
When people first get a chance to affirm something they value about themselves, they become less defensive about threatening information. The effect is well studied, though it depends on timing and context, so treat it as a nudge rather than a fix.
Cohen and Sherman, “The psychology of change”, Annual Review of Psychology, 2014
Inviting people to find something out (“I was surprised by this, were you?”) taps the one trait Kahan found that reduces biased reading of evidence.
Kahan's central advice: make sure accepting the evidence doesn't require someone to turn against their group. Offer a way to agree with the facts that still lets them be who they are.
Small per use, large in total, and paid by people other than the user. Figures below are the most credible recent ones we found, with their dates. Company figures are self-reported and not independently checked.
Google compares it to about nine seconds of television. Google reported 0.24 Wh for a median Gemini text prompt (Aug 2025); Epoch AI estimated about 0.3 Wh for a typical ChatGPT query (Feb 2025). Long or “reasoning” answers use much more.
Image models tested in 2023 averaged around 2.9 Wh per image, roughly ten times a text prompt; the heaviest used close to a phone charge. Newer models vary widely.
About 1.5% of world electricity in 2024, heading to just under 3% by 2030 in the base case (range about 800 to over 1,260 TWh). AI is the main driver of the growth.
US data centres used 176 TWh in 2023 (4.4% of the country's power) and could use 325–580 TWh by 2028.
Google puts a median text prompt at about five drops. In total, US data centres used roughly 66 billion litres directly for cooling in 2023, with far more used by the power plants supplying them. One academic study projects global AI water withdrawal of 4.2–6.6 billion cubic metres in 2027. Local impact depends on where the data centre sits.
Google, Aug 2025 · LBNL, Dec 2024 · Li, Yang, Islam and Ren, revised Mar 2025
Microsoft's emissions jumped 25% in its 2025 financial year as it built data centres for AI, with double-digit rises at Amazon and Google too. Google reported lower data-centre emissions for 2024, but its data-centre electricity use rose 27% that year; the cut comes from how clean-energy purchases are counted. Per prompt, Google reports about 0.03 g of CO₂.
On the largest US grid (PJM, 13 states and DC), capacity prices jumped from about $29 to $270–329 per megawatt-day between 2024 and 2026. The grid's market monitor attributes much of that to expected data-centre demand; residential bills in parts of Maryland and Ohio rose by roughly $16–18 a month. Others argue market design is as much to blame.
The record settlement approved in July 2026 between authors and Anthropic, about $3,000 per book, over pirated copies used to build training sets. The US Copyright Office (May 2025) also flagged “market dilution”: AI output crowding out the human work it learned from.
Around 5% of ChatGPT's users paid in late 2025, and OpenAI had an operating loss of about $8 billion in the first half of 2025, both according to the Financial Times. Free use is covered by investors for now; history suggests the cost later shows up as ads, data use or higher prices.
Research pack · first version 29 Oct 2026. Figures are the latest credible ones found on that date; per-use energy estimates are falling quickly, so check the date before quoting.
Quotes are kept short; everything else is summarised in our own words.