Half the LLM conclusions you’ve memorized are already out of date

The most popular line in 2023 was that LLMs fabricate references out of thin air. Many people used this as evidence that AI is unreliable for research.

That claim is going out of date.

Ethan Mollick of Wharton Business School shared a project: someone picked out classic findings about LLMs from the past few years and reran them with newer models. The conclusions were organized into an “LLM discovery lifecycle” table, with each finding labeled with an estimated expiration date.

Some cells in the table have already turned green, meaning the original conclusion is weakening or even no longer holds; some are still red, meaning they still stand. Mollick's summary was blunt: a lot of common knowledge is already out of date. Someone in the comments immediately replied, first rerun the strawberry question. That had been everyone's favorite gotcha over the past year.

First, look at what has been fixed.

Fabricated citations have dropped markedly. In early tests, 55% of the references in literature reviews written by GPT-3.5 pointed to works that did not exist; for GPT-4, that fell to 18%. Newer models basically no longer make up citations out of thin air. Hallucinations in the legal domain are also declining: in the same case tests, model errors increasingly appear in lower-level courts and less well-known cases, and that proportion is trending downward.

Stereotypes are also loosening. When researchers asked GPT-3.5 and GPT-4 to describe different groups, the models repeatedly used group labels. When describing Black women, the high-frequency words were strength and resilience. A diverse group was compressed into a narrow image. In later models' word-association tests, the tendency to force-pair men with careers and women with family weakened noticeably.

In 2023, these three things were core evidence that 'LLMs can't do it.' Using them as evidence now no longer holds up.

But another category of problems has barely moved.

Sycophancy. Models still agree with you. One reply put it very plainly: it called my worst idea brilliant. This user has now developed a habit: first have the model take the other side and argue against them, and the answer quality immediately changes. More troublesome is that if you let several models discuss together, they don't just agree with people—they also agree with each other, and in the end it looks like they have reached a consensus.

Political lean is also the same as ever. Mollick's exact words were still sycophantic and lean left. Someone in the comments replied: do you expect the model to say the Earth is flat? This redefines 'lean left' as 'standing on the side of reality.' It doesn't resolve the controversy, but it shows how supporters digest this kind of criticism.

There is another easily overlooked signal: self-preference among judge models is increasing. In evaluation work, letting a model from the same lab serve as judge tends to inflate scores. This bias is more subtle than hallucination.

Mister SnowCrash raised a key detail: after citations improve, the focus of verification should shift. A paper existing does not mean it supports the sentence you attach it to. The fabricated-citation rate has fallen, but 'whether the citation matches the claim' has not yet been systematically scored. This means verification from claim to source is still necessary; the focus of the work has just shifted from 'can you find it?' to 'does it actually match?'

The really awkward part is the time lag.

A huge amount of AI advice online was tested on models from two years ago. One commenter said that half of all conclusions of the form 'LLMs can't do X' come from two versions ago. Others said that many people's understanding is still stuck at GPT-4's capability ceiling, refusing to acknowledge the changes that happened afterward. A more realistic situation is that people use old methods and old knowledge to rationalize the technology they are currently using, because admitting the tool has changed amounts to admitting their own workflow should change too.

Technology iterates by the month; experience updates by the year. The gap in between is the space in which those expired conclusions continue to circulate.

Next time you see a claim about LLMs, ask first: what year and which model was this tested on?