AI Can Search the Web. It Still Doesn’t Know What to Look For.

An essay by an unnamed author argues that AI systems such as GPT can search the web and hold large context windows but still fail at the upstream problem of deciding what to retrieve before making a judgment, citing Liu and colleagues' "Lost in the Middle" finding that models can fail to use information depending on where it appears in a long context. The author illustrates the gap with a peer-review email analysis in which GPT's explanation changed each time it was directed to consult the submission timeline, saved documents, and normal review process, and points to Teyler and DiScenna's hippocampal memory indexing theory, Polyn and colleagues' free-recall study, and Badre and colleagues' semantic memory work as evidence that retrieval is cue-driven rather than capacity-driven.

Recently I was using GPT to work out where one of my academic papers stood in peer review. A status message had arrived from the editorial office, and I wanted to understand what it meant. I went through it with GPT. The first answer was careful. It read the wording of the email, avoided claiming to know the sender’s intentions, and kept several possibilities open. That isn’t the wrong posture. But something was missing. I told it to stop reading the email in isolation and look at the whole timeline of the submission. Then I told it to compare against how peer review normally runs. Then I pointed it to the documents I had saved and to the history of the revisions. Each time, its explanation changed. What matters here is not which reading turned out to be right. What I noticed sat further upstream. GPT had a lot of what it needed. It had the earlier conversation. It had the files. It could search the web if it chose to. And yet in its first answer, it didn’t really ask itself: what do I need to look at next before I can judge this? I wasn’t handing it conclusions. I was telling it where to look before it concluded. I think that gap matters more than it seems to in today’s AI. I’m not reproducing any private correspondence or review content here. The subject is the information-selection problem that surfaced while I worked through it with GPT. Memory has quickly become one of the big themes in AI: long context windows, persistent memory, project-level files, external databases, RAG. All of it is useful. But storing more does not, by itself, make judgments better. Liu and colleagues’ “Lost in the Middle” showed that even when the needed information sits inside a long context, models can fail to use it well depending on where it appears 1 . That result concerns the models and tasks studied at the time, not every model in use today. It is still a good reason to separate “can be put in the input” from “can be used when it’s needed.” In my case, too, the information existed. What was weak was the move of working backward from the current question to call up the relevant timeline, the saved documents, the normal process, and the right points of comparison. Owning a large library is different from knowing which book to open right now. Beyond capacity, the next problem for AI memory may be what drives retrieval. We don’t unfold our entire life’s memory into awareness every time we make a judgment either. The current situation acts as a cue, and related memories and knowledge get called up from there. Teyler and DiScenna’s hippocampal memory indexing theory treats the hippocampus as an index for re-accessing the distributed cortical representations that were active during an experience 2 . In that view, memory is less a box that holds each event whole and more a mechanism for reconstructing the needed patterns of activity later. In a free-recall study, Polyn and colleagues found that cortical activity corresponding to a category people, places, objects appeared several seconds before participants recalled and said aloud an item from that category 3 . In human memory search, the context and cues active at the moment are related to what gets pulled up next. Badre and colleagues, working with semantic memory tasks, separately examined the control involved in retrieving information from weak cues and the control involved in selecting what’s needed among competing candidates, and reported that the two corresponded to different activity in the ventrolateral prefrontal cortex 4 . There’s no need to copy brain regions over to AI. What’s interesting is how the functions are divided. The current object becomes the cue. Related past material gets called up. Among the candidates, the ones that matter this time get selected. If that still isn’t enough, you go looking for the next piece. AI memory may need more than topic-based archives like “work,” “family,” and “papers.” If memories carried return paths, such as which assumption I got wrong last time, what I checked that changed the judgment, and under what conditions an old conclusion stops holding, then memory could work less like a warehouse for replaying the past and more like an index for correcting present reasoning. Search systems usually look for information related to the question. In reasoning, related isn’t always enough. If explanation A and explanation B are competing, what you really want is information that can change which one survives. Something that resembles both won’t do that. Markant and Gureckis compared active learning, where learners choose their own examples, with a condition in which the same examples are handed to learners passively 5 . In their analysis, what mattered was not only the act of choosing but that learners could pick the next example according to the hypotheses they currently held. With causal inference the stakes rise. Steyvers and colleagues showed that learning a causal structure improved when participants could choose their own interventions and observe the results, compared with observation alone 6 . Deciding which intervention would tell competing causal hypotheses apart becomes part of the problem itself. A web search is not an experimental intervention, and piling up search results doesn’t prove causation. The design question carries over all the same: What information could change the current explanation, rather than merely reinforce it? I think this is one of the lines between an AI that can search and an AI that can investigate. Look at the AI side, and something interesting is happening. Modern LLMs carry knowledge and patterns acquired through large-scale pretraining. Petroni and colleagues examined how much relational factual knowledge can be drawn out of pretrained language models and showed that knowledge captured in the parameters can be queried 7 . That doesn’t make a language model a complete or accurate database. Combining that internal knowledge with external retrieval is the basic direction of RAG. Lewis and colleagues combined a pretrained model’s parametric memory with a retrievable, non-parametric memory 8 . Now add long-term memory, user-specific projects, files, web search, and a range of tools. At that point, AI can become more than a system that answers from what it already knows. It can think of a field to compare against from general knowledge. It can go back to an earlier correction through its own memory. It can pull the original wording from a file. If current information is needed, it can go to the web. If the current state only lives in an external service, it can check with a permitted tool. I’m not simply saying this beats human memory. People have bodies, experience, purposes, meaning-making, and an ongoing relationship with the world around them. AI has a different kind of strength: It can use broad learned knowledge as a cue, then go out into the world on the spot to fetch what it’s missing. With that combination, the bottleneck may shift from how much was stored to what to go and get next. If I wrote here that AI research hasn’t noticed this idea, this article would fail its own test. Research that alternates between reasoning and retrieval is already well along. ReAct interleaved a language model’s reasoning with actions in an external environment. The model gains information through actions and uses the results in later problem solving 9 . IRCoT is even closer to what I’m describing. In multi-step knowledge questions, what to retrieve depends on the reasoning so far, and that reasoning depends on what was retrieved earlier. Trivedi and colleagues built on this interdependence by interleaving retrieval with chain-of-thought reasoning 10 . Self-RAG introduced a mechanism for the model to assess when to retrieve, whether retrieved passages are relevant, and whether its output is supported by them 11 . Search-R1 uses reinforcement learning to train reasoning that includes multiple rounds of interaction with a search engine 12 . In 2025, Geng and colleagues studied tasks that reverse-engineer black-box programs and formulas 13 . They compared an active condition, in which the model chose its own inputs and observed the outputs, with a condition in which it passively received examples. Performance improved in the active condition, and passing the same data passively to a separate run did not produce the same performance. These are controlled tasks that return reliable input-output pairs, and the conditions differ in more than who does the choosing. Still, the result makes a case for evaluating the process of selecting information while updating hypotheses, alongside the information collected. So I’m not claiming to have invented a new search algorithm. What bothers me is a more practical gap. Search, memory, files, and tools can all be available. If the user still has to keep saying “now look at this,” an important part of information selection is still being done by the human. There is still a distance between what’s possible in research and what fires spontaneously, at the moment it’s needed, in an actual conversation. Working backward from this failure, the following flow seems natural to me: Current Object → Initial Causal Model → Missing Evidence → Source Routing → Evidence Typing → Causal Update → Commit or Continue First, pin down what you’re actually trying to judge. Next, build several candidate explanations from the current information and learned knowledge instead of betting on one. Then ask: What would change the ranking of these candidates? Only then decide where to go. If past exchanges or corrections are needed, go back to memory or the project. If you need the fine detail of the original material, read the files. If current rules, research, or product information may have changed, search the web. If the current state only exists in an external service, use a tool. Whatever can’t be obtained stays marked as unknown, or goes back to the human as a question. After retrieving, keep track of where each piece came from. General knowledge that surfaces from pretraining is a strong prior for deciding what to look into. On its own, it proves nothing about this particular case. Memory can bring back the past, but it can’t guarantee that an old state still holds. Secondary sources on the web help with comparison, but they don’t carry the weight of primary documents. Then update the causal model with the retrieved evidence. This is the important part. Search is there to test the first answer as much as to support it. New evidence is allowed to flip the ranking of explanations. And if the question was framed wrongly to begin with, the causal model itself can be rebuilt. Finally, if missing evidence that could still change the judgment remains, keep going. When further retrieval is unlikely to change much, answer, and say what uncertainty remains. These seven functions don’t need to be exposed to the user as a step-by-step trace. What matters is that the system performs them when the task requires them. This also suggests that caution should act at the point of commitment, not by weakening the reasoning from the start. Build strong candidates from the available material, and then be strict about what gets committed as fact. If you can’t know what someone was thinking, mark only their intentions as unknown. If you haven’t confirmed the current state, mark only that as unconfirmed. Uncertainty on the outside doesn’t justify thinning out the causality you can observe on the inside. Thinking boldly and asserting boldly are two different things. There’s an opposite failure. Being able to search doesn’t mean heading to the web every time a question comes up. In a decision task where participants could pay for information, Kobayashi and Hsu showed that the value of information includes its instrumental value, which comes from improving choices, and also the value of knowing an outcome in advance 14 . They also examined neural signals related to that value in the striatum and ventromedial prefrontal cortex. People sometimes gather information because they want to know, not because it helps. AI can fall into the same trap if the number of searches becomes the measure of performance. Collect ten articles that agree with your explanation, and if they all trace back to a single source, you have one piece of evidence, not ten. If your search terms are phrased to confirm your hypothesis from the start, the world you find will be skewed as well. So before searching, I think this is the question to ask: What would change my current judgment? And after searching, one more: Is what I just retrieved evidence for that question, or just something on the same topic? When those questions stop getting answers, the volume of searching may be rising while the reasoning gets no deeper. Evaluating this idea doesn’t have to be complicated. Scatter the important information across several places: some in memory, some in files, some in current web information. Mix in outdated information and irrelevant material. The user doesn’t say which source to look at. Then see whether the AI reaches the sources it needs on its own. The number of searches is one thing to measure, and only one. Did it find the needed material before a human pointed to it? Did it put current evidence ahead of outdated memory? Did it choose information that separates competing explanations? Did new evidence actually update its judgment? Did it avoid unnecessary searching? And how many times did a human have to step in with “now look at this” before it reached the same quality? This one episode is not an experiment proving the method works. But it did make clear, at least to me, what I was frustrated about. What I wanted is easy to say and hard to build: an AI that notices what its own reasoning is missing, and chooses the next source to fill that gap before I have to ask. The web offers a way back to the present world. Pretrained weights offer an enormous space of things worth checking. AI can already reach a great many places. Binding all of that into one line of reasoning takes one more thing. Memory gives an AI somewhere to return. Reasoning must decide when to return — and where. 1 N. F. Liu et al. , “Lost in the middle: How language models use long contexts,” Trans. Assoc. Comput. Linguistics , vol. 12, pp. 157–173, 2024, doi: 10.1162/tacl a 00638. Online . Available: https://aclanthology.org/2024.tacl-1.9/ https://aclanthology.org/2024.tacl-1.9/ 2 T. J. Teyler and P. DiScenna, “The hippocampal memory indexing theory,” Behav. Neurosci. , vol. 100, no. 2, pp. 147–154, 1986, doi: 10.1037/0735–7044.100.2.147. Online . Available: https://pubmed.ncbi.nlm.nih.gov/3008780/ https://pubmed.ncbi.nlm.nih.gov/3008780/ 3 S. M. Polyn, V. S. Natu, J. D. Cohen, and K. A. Norman, “Category-specific cortical activity precedes retrieval during memory search,” Science , vol. 310, no. 5756, pp. 1963–1966, 2005, doi: 10.1126/science.1117645. 4 D. Badre et al. , “Dissociable controlled retrieval and generalized selection mechanisms in ventrolateral prefrontal cortex,” Neuron , vol. 47, no. 6, pp. 907–918, 2005, doi: 10.1016/j.neuron.2005.07.023. 5 D. B. Markant and T. M. Gureckis, “Is it better to select or to receive? Learning via active and passive hypothesis testing,” J. Exp. Psychol.: Gen. , vol. 143, no. 1, pp. 94–122, 2014, doi: 10.1037/a0032108. 6 M. Steyvers, J. B. Tenenbaum, E.-J. Wagenmakers, and B. Blum, “Inferring causal networks from observations and interventions,” Cogn. Sci. , vol. 27, no. 3, pp. 453–489, 2003, doi: 10.1207/s15516709cog2703 6. 7 F. Petroni et al. , “Language models as knowledge bases?” in Proc. EMNLP-IJCNLP , 2019, pp. 2463–2473, doi: 10.18653/v1/D19–1250. 8 P. Lewis et al. , “Retrieval-augmented generation for knowledge-intensive NLP tasks,” in Proc. NeurIPS , 2020. Online . Available: https://arxiv.org/abs/2005.11401 https://arxiv.org/abs/2005.11401 9 S. Yao et al. , “ReAct: Synergizing reasoning and acting in language models,” in Proc. ICLR , 2023. Online . Available: https://arxiv.org/abs/2210.03629 https://arxiv.org/abs/2210.03629 10 H. Trivedi, N. Balasubramanian, T. Khot, and A. Sabharwal, “Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions,” in Proc. ACL , 2023, pp. 10014–10037, doi: 10.18653/v1/2023.acl-long.557. 11 A. Asai, Z. Wu, Y. Wang, A. Sil, and H. Hajishirzi, “Self-RAG: Learning to retrieve, generate, and critique through self-reflection,” in Proc. ICLR , 2024. Online . Available: https://arxiv.org/abs/2310.11511 https://arxiv.org/abs/2310.11511 12 B. Jin et al. , “Search-R1: Training LLMs to reason and leverage search engines with reinforcement learning,” arXiv:2503.09516, 2025. Online . Available: https://arxiv.org/abs/2503.09516 https://arxiv.org/abs/2503.09516 13 J. Geng, H. Chen, D. Arumugam, and T. L. Griffiths, “Are large language models reliable AI scientists? Assessing reverse-engineering of black-box systems,” arXiv:2505.17968, 2025. Online . Available: https://arxiv.org/abs/2505.17968 https://arxiv.org/abs/2505.17968 14 K. Kobayashi and M. Hsu, “Common neural code for reward and information value,” Proc. Natl. Acad. Sci. USA , vol. 116, no. 26, pp. 13061–13066, 2019, doi: 10.1073/pnas.1820145116. How this was written: The problem came out of a real conversation between me and GPT. GPT assisted with the literature search, organizing the argument, structuring, drafting the Japanese version, and checking sources; a source ledger records how much of each reference was actually read. Claude Anthropic reviewed the Japanese drafts independently and wrote this English version from my final Japanese text. I reviewed the English and made the final editorial decisions, and I am responsible for what is published. The parallels to human cognition are functional design analogies, not claims that brains and AI run on the same mechanisms. The evaluation proposed above has not been run. AI Can Search the Web. It Still Doesn’t Know What to Look For. https://pub.towardsai.net/ai-can-search-the-web-it-still-doesnt-know-what-to-look-for-1429d3cdb65f was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.