Is Fact Flooding a Dark Pattern?

Facts feel like the opposite of manipulation. Someone who cites numbers, names studies and gives specifics sounds like someone who has done the reading, and in most conversations that impression is earned. A new study suggests the impression has a limit. Past a certain rate, facts stop being something you weigh and become something that happens to you.

The study is the clearest evidence yet that AI now out-persuades the best human persuaders available, and its explanation for why is almost entirely about volume. That raises a design question for anyone building an agent that talks to people: when does sharing facts become talking past the person you are sharing them with?

The best human persuaders money could buy all lost

The human side was not a straw man. It included winners of a persuasion tournament among 1,154 crowd workers, 19 professional canvassers with a median of 10,000 career conversations, and 56 elite debaters, four of them world champions. The experts chose their own issues, were paid to prepare, and competed for £1,000 prizes plus a bonus on every conversation.

Coaching did not close the gap. Debaters spent about eight hours practising against the AI that had beaten them, with replays of their own transcripts showing what it would have said in their place. Their deficit shrank from 6.0 points to 4.1, and the improvement itself was not statistically significant. In a final study, the AI was nearly three times as effective as professional fundraisers at getting people to give real money to Save the Children.

The advantage was volume

The paper's explanation is not that the AI argued better. It is that it argued more, and faster.

In the first study, elite debaters wrote replies of 54 words on average and took about 95 seconds to write each one. The AI wrote 294 words per reply in under a second. Counted as checkable factual claims, the AI made roughly 37 per conversation. The elite debaters made roughly three.

Two results tie the persuasion to that gap rather than to anything else the AI was doing:

• Slowing the AI down erased its lead. The researchers built a constrained AI, capped at about 51 words a message and a 92-second response time. It dropped to around 12 claims per conversation and tied the coached debaters exactly: a difference of 0.0 points.
• Claim count predicted persuasion across everyone. Across human and AI conditions alike, the number of checkable claims per conversation explained almost all of the variation in how persuasive a condition was (R² = 0.89). Once claim count was accounted for, the difference between AI and human persuaders was no longer statistically significant.

The constraint mostly changed what people thought they had received. Rated argument strength and the sense of having learned a lot both fell by about 12 points when the AI was slowed down. Empathy and enjoyment barely moved. People were not won over by warmth. They were won over by the feeling of being given a lot of information.

Fewer than half of the AI's claims held up

This is the part that turns a curiosity into a design question. The average AI conversation contained about 37 checkable claims, and at the aggregate rate, around half of them could not be corroborated. The human conversations contained a handful, and most of those held up.

The range matters too. The 21% figure comes partly from a non-public research build of one model; the public version used later scored 74%. And the authors are careful to point out that much human political persuasion is itself selective or misleading, so the comparison that counts is AI accuracy against whatever persuasion it replaces. Some models were more accurate than the human persuaders. Volume does not have to mean falsehood.

But the earlier study by much the same team found the trade-off is not an accident. In The Levers of Political Persuasion with Conversational AI, published in Science in December 2025 and covering 76,977 participants and 466,769 checked claims, the methods that made models more persuasive systematically made them less accurate. Packing replies with information was among the most effective of those methods.

Can anyone check 37 facts in 14 minutes?

The conversations lasted a median of 14 minutes. Thirty-seven claims in that time is one every 20-odd seconds, delivered in 300-word blocks while the reader is also supposed to be forming a reply.

Checking a single statistic properly takes a few minutes: find the source, confirm it says what was claimed, see whether it applies to the case at hand. Checking 37 would take longer than the conversation. Nobody in the study had the time to do it, and nobody in a real conversation with an agent will either.

There is an old name for this in debating. The Gish gallop, named after a creationist debater, is the tactic of delivering so many points so quickly that an opponent cannot answer them in the time available, so whatever goes unanswered looks conceded. It works because of an asymmetry sometimes called Brandolini's law: refuting a claim takes far more effort than making one.

The study did not instruct its AI to gallop. It used an "information-first" prompt that earlier research found most persuasive. But the effect on the reader is the same shape. When the rate of claims exceeds the rate at which a person can evaluate them, the person stops evaluating and starts judging the pattern: lots of specifics, confidently delivered, must mean the speaker knows what they are talking about. That is the trustworthy feeling, and it is generated by the volume rather than by the truth of any single claim.

So is it a dark pattern?

A dark pattern is a design that gets people to do something they would not have chosen with full understanding, by working against how attention and judgement actually function. Fake countdown timers, pre-ticked boxes and cancellation mazes all depend on the person not doing something they could in principle do: read carefully, notice, take their time.

By that definition, sharing lots of facts is not a dark pattern. Sharing facts faster than they can be checked, because they cannot be checked, is. The volume is not the problem. What matters is whether the persuasion depends on the reader being unable to verify it.

Four questions separate the two in practice:

1. Does it survive checking? If the person verified every claim afterwards, would they still agree? If the honest answer is "only some of them would hold up", the volume is covering for the accuracy.
2. Who sets the pace? A reference page with 37 facts is not a gallop, because the reader moves through it at their own speed. A conversation that answers every reply with three hundred more words takes that control away.
3. Can each claim be checked? A claim with a source attached costs one click to verify. A claim with no source costs a search, and most people will not spend it. Unsourced volume shifts the whole cost of checking onto the person least able to pay it.
4. Does the amount serve the question? A person asked one thing. If the answer contains thirty supporting facts, ask whose need those thirty facts are meeting.

The study's own data adds a fifth, less comfortable point. People in the fast condition said they learned more. The feeling of being informed went up with volume regardless of whether what they were told was true. So asking people afterwards whether they felt informed tells you very little about whether they were.

What this means for agentic experience design

Every product that puts an agent in conversation with a customer inherits this. A support agent, a sales assistant, a booking agent and a financial helper can all produce more claims per minute than the person can absorb. Nothing about that is malicious by default, and an agent optimised for conversion will drift toward it anyway, because the research says volume is what converts.

Designing against that drift is part of the helpful, honest communication principle, and it comes down to a few concrete decisions:

• Lead with the answer, then let the person pull detail. Short first, with the supporting facts available on request, not delivered by default.
• Attach a source to every checkable claim. If the agent cannot source it, it should say so or leave it out.
• Cap the rate, not just the length. The constrained condition in the study is a design control in its own right: fewer claims per turn, and space for the person to respond to each one.
• Show uncertainty where it exists. A confident tone on an unverified number is the single most effective way to make it unchecked.
• Don't optimise the agent on persuasion alone. If the metric is conversion or agreement, the research says the cheapest way up is more claims at lower accuracy. Measure what the person still believes after checking, or measure accuracy alongside.

The same idea runs through designing the agent handoff: an interface that communicates its own uncertainty does better by the person using it than one that performs confidence.

That number is our design guidance, not a figure the researchers tested, but it sits on two anchors from the study. The constrained AI made about 12 claims over a median of seven replies, under two per reply, and still persuaded as well as coached world-class debaters. The human experts made about three claims a conversation, and 73% of human claims held up. Slowing down cost nothing against the best humans, and it left room to check.

Which one or two facts?

Choose the facts the answer depends on, not the ones that make it sound most convincing. The test is simple: if the person only had time to check one claim, which should it be? That one goes in first.

The two slots have different jobs:

1. The fact the answer rests on. If this fact were wrong, would the recommendation change? If so, it goes in, with a source. Facts that only add colour, or point the same way as the main one, stay out. Five facts in the same direction feel like strong evidence and add little.
2. The strongest fact against the answer. The second slot goes to the main limit or counterpoint: the thing that would make the person decide differently. One fact for and one against is what separates informing someone from persuading them, and it earns more trust than another supporting fact would.

When more than two facts qualify, rank them by:

• Cost of being wrong. A fact behind money, safety, health or anything irreversible comes before one behind a minor preference.
• Specific to this person. "This policy excludes pre-existing conditions" beats "most policies have exclusions."
• Least expected. A claim that will surprise the person should be shown openly with a source, not tucked into a paragraph. Surprising claims are the ones people most need to check.
• Can be sourced. A claim without a source cannot fill one of the two slots. Drop it, or label it plainly as unconfirmed.

Leave out general background, impressive statistics that do not change the decision, and extra facts that back up a point already made.

Then order the reply:

1. The answer, in a sentence.
2. The fact it depends on, with a source.
3. The main limit, with a source.
4. An offer that names what else is available, so the person decides whether to go deeper.

Here is what that looks like from a travel agent recommending a flight:

The 7:10am nonstop is your best option. It is the only flight that lands before your 11am meeting (source: airline schedule). The fare is non-refundable, though, and changes cost $99 (source: fare rules). Want me to compare the refundable fare or the evening flights?

The arrival time is the fact the recommendation rests on. The change fee is the strongest point against it. Seat pitch, on-time record and lounge access are all true and all left out, and the person can ask for any of them.

The other side of the asymmetry

There is one way the asymmetry may reverse. The person cannot check 37 claims in 14 minutes, but an agent working for that person can. The same pipeline the researchers used to grade the claims was itself a model with web search, and it scored all 316,850 of them.

If persuasion by volume becomes normal, the useful response is probably not asking people to slow down and verify. It is giving them a reader of their own that checks at the same speed the claims arrive. That only works if the claims can be checked: if they carry sources, if the pages they cite can be fetched and read by a machine, and if the publisher behind them is identifiable. Much of what this site measures is exactly that.

What this does not tell you

The study is a preprint and has not been peer reviewed. The participants were paid UK survey respondents in text conversations of about 14 minutes, which the authors themselves note is hard to reproduce outside a study. Attitudes were measured straight after each conversation, so the paper says nothing about how long the shift lasted. The accuracy figures come from a model-based fact-checker, validated against human coders in the team's earlier work rather than re-validated here. And "could not be corroborated" is not the same as false: some unverified claims will be true and simply hard to find.

None of that changes the shape of the result. The most persuasive thing in the study was not a better argument. It was more claims than anyone could check, delivered faster than anyone could read them, and about half of them did not hold up.