LLMs Are Time Machines Without Distance Tracking

Large language models function as time machines that compress the journey from inquiry to decision, but they strip away the critical metadata of that travel. An LLM takes your present self, armed with a question, and returns your future self, equipped with an answer. This compression is real and profound. In the past, answering a serious question required a library visit. You consulted books, took notes, followed references, and spent days building enough context to make a decision. Search engines compressed this timeline into hours. You queried, received results, and traced clues through those answers until you had a vivid enough picture to act. Answer engines now compress the process into seconds. The starting point and destination remain the same, but the travel is nearly nonexistent.

Every limitation of previous technologies persists in this new format. If a library lacked the right books, you were stuck. Google often surfaced sources without the editorial oversight that publishing a book required, allowing errors and confident nonsense to scale beyond human sorting capabilities. Yet, people decided anyway. The speed is novel; the underlying uncertainty is not. What is truly new, and what concerns professional publishers, is the loss of path metadata.

The research journey was never just about moving toward an answer. It simultaneously signaled the value of that answer. Encountering only three books and no journal articles indicated a thin topic. Contradictory sources suggested controversy. A search yielding nothing meant uncharted territory. A four-hour research session felt fundamentally different at its conclusion than a four-minute one. That difference was information. People did not consciously read these signals, but they were present in the record, shaping how firmly individuals committed to their findings. Path metadata is a byproduct of travel rather than a designed feature. It does not survive into the destination, yet until recently, reaching the destination required generating it along the way.

An answer engine delivers a conclusion in confident prose regardless of whether the evidence behind it was deep or absent. The compression is lossy because it strips out the exact signals used to evaluate the output. Users arrive at answers without the means to judge them.

Empirical Evidence of Cognitive Deficit

For years, this phenomenon was arguable but not demonstrable. That changed with research published in October 2025. Shiri Melumad and Jin Ho Yun, marketing professors at Wharton, conducted seven experiments involving 10,462 participants and published their findings in PNAS Nexus. Participants learned about common topics, such as planting a vegetable garden or spotting financial scams, either via an AI summary or standard Google links. They then wrote advice for others based on their learning.

Those who used AI came away knowing less, even when both groups received identical facts. They engaged less with the material, and their subsequent advice was sparser, less original, and less likely to be adopted by recipients. Crucially, when the model provided live web links alongside its answer, participants did not click them. Once the summary arrived, the adjacent sources lost their appeal.

Pew Research Center observed similar behavior in natural browsing settings. Tracking 900 U. S. adults across 68,879 Google searches in March 2025, Pew found that when an AI summary appeared, users clicked normal search results on only 8% of visits, compared to 15% without summaries. Clicks on cited sources within summaries occurred roughly 1% of the time. Users ended browsing sessions entirely on 26% of pages with summaries, versus 16% without. While Pew notes this is association rather than causation and limited to one month and US Google users, the convergence with controlled experiments is significant.

This aligns with earlier work. In 2015, Yale researchers found that internet searching inflated perceived knowledge, confusing access with understanding. This effect persisted even after searches yielded no results. This complicates the argument that the old journey taught good judgment, as it did not. The journey left behind friction, elapsed time, and visible source variety. Melumad and Yun demonstrate that removing these elements creates a deficit in user output. Additional studies from Microsoft Research and Carnegie Mellon in 2025, alongside Dirk Lewandowski’s 2026 study on information regret, reinforce that greater confidence in AI predicts less critical thinking.

The Loss of the Free Correction Mechanism

This shift transforms the issue from user behavior to business impact. Under the old model, thin or wrong answers were survivable because the journey repaired them. A reader might encounter a bad summary, continue searching, and land on the correct page, effectively swapping the error for accurate content. This acted as the immune system of the information economy. It cost nothing and happened millions of times daily.

With source-click rates dropping to approximately 1%, this correction mechanism fails. The time machine skips not only the journey but also the repairs that occurred along the way. Misrepresentations that were once temporary now persist. The asymmetry is costly. The old repair was automatic and free, driven by curiosity. The replacement requires payment. Publishers must produce evidence, wait for crawling and retrieval, and hope it is weighted correctly, with no control over timing or confirmation of success.

Strategic Implications for Content Placement

Treating AI citation as a referral channel is a misvaluation. The value lies in being inside the answer that drives action, not in the trickle of traffic that escapes. Consequently, inbound leads are not uninformed; they are confidently underinformed. Content written to educate a curious researcher fails to engage someone who believes the research is complete.

A decade of content strategy relied on a staircase: definitional explainers at the top, comparisons in the middle, and depth at the bottom. The top of this staircase now occurs before users reach your site. Those who do arrive are further along the decision journey but not better informed. They carry the confidence of thorough research alongside the actual depth of a single paragraph read. Beginner content talks down to them, ending visits quickly. Advanced content assumes a vocabulary they can repeat but have not earned, ending visits politely.

Deleting beginner content is not the solution, as it still feeds model training. The change must be in placement. The content that models cannot replicate must now serve as the front door, replacing the basics that happen elsewhere. Pages that are clean, structured, and explanatory are easiest for models to absorb, meaning the pages built to greet newcomers are no longer needed by newcomers. You should stop expecting your existing content stack to greet leads who have been fast-forwarded past its opening.

Internal Reflection and Systemic Blindness

Finally, this dynamic reflects back on creators. You are also a user of these systems. Competitive analyses, strategy decks, and recommendations derived from model synthesis exhibit the sparse, less original output measured by Melumad and Yun. You may feel more certain about these outputs than the process warrants. The recipients of such advice find it thin but cannot articulate why.

The time machine works, taking you from question to decision in seconds. Most decisions are acceptable. However, it deposits you somewhere without indicating how far you traveled or what you bypassed. Understanding the system’s operation is essential not because the answers are necessarily wrong, but because you cannot tell if they are.