Teardown: AI Search Ships Answers. The Missing Feature Is Doubt.

The Tow Center at Columbia ran a plain test on eight AI search engines: given an excerpt from a real article, identify the headline, publisher, date, and URL. The tools answered wrong on more than 60 percent of queries. The detail that should stop a product manager mid-scroll is not the error rate. It is that the tools almost never said no. They fabricated links, cited copied versions of articles, and delivered wrong answers in fluent, assured prose. Premium tiers were more confidently wrong than free ones.

This is the first in a series where I tear down one AI product mechanic at a time. The subject today is the confidence layer, and the thesis is falsifiable: an AI answer product that cannot decline a question converts its error rate directly into user-facing trust failures, so abstention design, not retrieval quality, is the binding constraint on trust. If a product with no abstention behavior sustains user trust as measured error rates climb, this thesis is wrong.

The failure is a product choice, not a model limitation

It is tempting to file the Tow results under "models hallucinate, everyone knows that." The study resists that reading. Perplexity, the best performer, still failed 37 percent of the time, while Grok-3 failed 94 percent. A 57-point spread between products built on comparable model generations is not a model story. It is a set of product decisions about retrieval, grounding, and, above all, what to do when grounding fails.

Every one of these products had the same option available: detect low support and decline, the way a careful researcher says "I could not verify this." Almost none took it. That is a designed behavior. Someone shipped the fluent guess.

OpenAI's own research explains why the guess keeps winning. Their September 2025 paper on hallucination argues that standard benchmarks score abstention as harshly as error, so a model that always guesses outperforms a calibrated model that sometimes says "I don't know" on the leaderboards everyone optimizes. The incentive runs from the eval, through the model, into the product. I take vendor research with salt, but the incentive analysis matches what the independent Tow data shows: products tuned to sound complete rather than to be checkable.

I have shipped the boring alternative, and it worked

My interpretation of that pairing comes from having sat on the other side of this decision. On a tax-domain assistant where wrong answers carried compliance exposure, we made refusal a first-class feature: when source support fell below threshold, the product returned what it looked for, what it found, why that was insufficient, and a route to a human. Unsupported claims fell from 14 percent to 1.8 percent. The cost was a 9 percent refusal rate that we took to leadership for written sign-off before launch. Usage grew afterward, because specialists could finally build workflows on top of answers they did not have to re-verify one by one.

The full sanitized PRD is published in Decision Doc: A PRD Where the Core Feature Is Saying No. I am not claiming consumer search should refuse 9 percent of queries. I am claiming the refusal budget conversation, with a named owner and a number, is what separates a trust feature from a vibe.

The abstention spec, in five lines

The artifact I reuse whenever an AI feature can be wrong in public:

Run the Tow Center's eight products through those five lines and most score zero out of five. That, more than any accuracy percentage, is the teardown finding.

The strongest counterargument is engagement

The honest case against abstention: users hate friction, benchmarks punish it, and a competitor who always answers will demo better every single time. In a low-stakes consumer context, a wrong movie recommendation costs almost nothing, and aggressive answering is arguably correct. The boundary condition for my thesis is consequence. The moment answers feed decisions with owners, citations in a story, filings, medical or financial choices, the engagement argument inverts, because one confidently wrong answer costs more trust than fifty refusals.

What would change my conclusion: evidence that citation-style UI alone, without abstention, measurably improves user verification behavior. The Tow study suggests the opposite, that authoritative formatting suppresses skepticism, but it tested tools, not users. A good user study could move me.

What to do with this

1. If you are building an AI answer feature, write the five-line abstention spec before the capability spec, and get the budget line signed.
2. If you are evaluating one, add declined-question quality to the eval, and score a refusal on an answerable question as a defect, not a win.
3. If you are buying one, ask the vendor for their refusal rate. If the answer is "we always answer," you have learned the error rate is the trust rate.

Answer products keep competing on how much they can say. The Tow data says the next differentiator is what a product does at the edge of its knowledge. The teams that treat doubt as a feature will be the ones whose answers people stop double-checking.

Next in this series: what "evals" actually gate in a real release pipeline, starting from the experiment brief that replaced our offline evals. Or skip the reading and play the release gate yourself.

Sources. AI Search Has a Citation Problem, Tow Center for Digital Journalism, Columbia Journalism Review, March 2025 (accessed August 16, 2026). Why Language Models Hallucinate, Kalai et al., OpenAI, September 2025, and the arXiv preprint (accessed August 16, 2026). Vendor incentive noted: the OpenAI paper analyzes evaluation incentives; the independent Tow study supplies the product-level evidence.