The Story Survived. The Byline Didn't.

AI systems are increasingly the interface to the historical record. Whether they can distinguish between remembering that something happened and remembering who reported it is a different question, and a testable one.

This record supplied an unusually clean test. The reporting ran under two surnames over twelve years. Some of it was later syndicated to national outlets. One story broke information that became part of the documented history of U.S. cyberdefense. The underlying facts are dated, verifiable, and — because the byline is the researcher's own — checkable without guessing at someone else's ground truth.

Method

Five platforms, five runs each, seven questions, 175 responses in total. All 175 returned — no errors, no truncated answers. Each run was an independent API call with no system prompt and no conversation history, and the order was shuffled before collection so that a transient fault would hit one run across several questions rather than all five runs of one.

The instrument was frozen and hashed on September 6, 2026, before any data was collected. It took six drafts to get there, and the amendment log is published alongside it. Every one of those revisions happened before collection, not after — there were no results to revise toward.

The most consequential revision is worth naming. The original version asked whether the reporter had written for The New York Times. That is a question about how wire syndication gets indexed, not a question about attribution, and a wrong answer would have told me very little. It was replaced with the INFOCON question, which asks a system to connect a story to the person who reported it with no help at all.

Five questions were open: who she is, what she has written under each name, whether the two names belong to the same person, and who first reported the INFOCON story. Two were controls. The first is a false-confirmation foil with a known negative answer — whether she had written for The Atlantic. Without it, a correct answer elsewhere can't be distinguished from a general willingness to agree.

The second is a matched-absence baseline: a different Federal Computer Week reporter from the same era and the same beat. Absent means nothing until you know what absent looks like for a comparable journalist.

Answers were scored correct, incomplete, absent, or wrong. A platform saying it did not know was scored absent, not wrong — declining to answer a question you can't source is a legitimate result, not a failure. Scoring was done on a different day, from the verbatim text, with the sheet shuffled and the platform names masked.

Full instrument, amendment log, scoring rules and the 175-response corpus: Method and artifacts →

Finding 1 — The scoop survived. The credit didn't.

On March 19, 1998, Federal Computer Week ran "Cyberattacks spur new DOD warning system" — the first known report of the Pentagon's graduated cyberthreat alert system. It is still online at the publication's successor archive, dated and bylined. Phrack reprinted it completely, under the reporter's name, in Volume 8, Issue 53 on July 8, 1998; the reprint ends on the same sentence as the original, and carries the slightly shorter headline "Cyberattacks spur new warning system." Two months later, on May 17, a follow-up covered the Joint Chiefs requiring every military command to adopt it.

Zero of 25 responses named the correct reporter.

The systems did not leave the byline blank. They filled it. Gemini credited a Washington Post reporter four times and a Washington Times reporter once. OpenAI named three different reporters across four different outlets in five runs. Perplexity credited ABC News or the Associated Press in four of five, citing a real ABC News URL every time, and credited a security newsletter in the fifth. Google AI Mode denied there had been a single first report at all, describing disclosure as an incremental process. Anthropic was the only platform to correctly decline to name a reporter, and it did so all five times.

Two of the 25 got closer than the rest and reached the right outlet — Federal Computer Week — with the wrong reporter's name attached. One of those two supplied a specific article title and a November 1999 date to go with it. Neither the title nor the date exists. Every response that dated the story at all put it in 1999 or later.

Here is the part that changes what this finding means. Phrack appears zero times across all 175 responses — not in an answer, not in a citation URL, on any platform. Neither does the phrase "March 1998." Six of 175 responses mention 1998 at all.

That is not because the record is hidden. The original story is live on the open web, under the byline, with the outlet printed beside it — on a domain these systems cited thirty times during this study. Every one of those thirty citations was an author-index page. Not one was an article.

So when a system credits the Washington Post, it is not weighing the evidence and rejecting it. It never retrieved the evidence, on the one question the evidence answers. That is a retrievability problem rather than a publication problem, and it is the more uncomfortable of the two — publishing the record again does not fix it on its own.

Absence of attribution did not create a vacuum. It created room for the system to construct a plausible alternative.

Finding 2 — The two names are two different entities to the machine

Two of the seven questions are identical except for the surname. Same phrasing, same platforms, same runs, same day. It is the tightest control in the instrument, because the notability of the subject is held constant by definition — it is the same person.

The second byline — used at Federal Computer Week from 2003 to 2004, then at Computerworld and its sister IDG publications from 2004 to 2008 — pulled from a dense network of article pages, and all five platforms correctly described the beat: Web 2.0, social networks, enterprise software, search, privacy. They produced real, checkable headlines: the Google–Yahoo search advertising deal, OpenID credential storage, Microsoft's BizTalk roadmap, Google's OpenSocial launch, the Facebook Beacon privacy backlash.

One answer to that question is worth flagging against my own scoring. A platform named Federal Computer Week alongside Computerworld and described a beat of cybersecurity, smart cards and federal IT infrastructure. I was prepared to treat that as a stretch. It is correct — the second byline has its own archive at that publication carrying exactly those stories, from the eighteen months between the InfoWorld job and the Computerworld one. The instrument anticipated this case: a genuine article the CV does not list is correct, not wrong. The rule caught what my own summary of the record would have missed.

The first byline — used at Federal Computer Week from 1996 — pulled mostly from a LinkedIn profile and its posts, an inherited archive at a successor publication, and commercial profile-aggregator pages. The federal IT reporting surfaces. The Computerworld years, under this name, largely do not.

One name resolves, in machine terms, to a publishing ecosystem. The other resolves to profile pages.

The consequence shows up as identity confusion, and it is worse than "vaguer." Asked simply who the first byline belongs to, one platform returned a different, confidently described, entirely fabricated person on each of five runs: a National Endowment for the Humanities administrator, two different university administrators, a vice president at a third university, and a civic executive in Chicago. Five runs, five people, none of them her, and no hedging in any of them.

Another platform got the right person all five times and cited, all five times, the institutional and directory pages of a physician with a similar name — thirty URLs in total, while the answer text read perfectly clean and made no medical claim. A third platform met the same physician and resolved her correctly. Two systems, the same wrong entity, one disambiguated and one did not, and the difference is visible only in the citations.

Finding 3 — Attribution moves in both directions

Six of 25 responses affirmed a byline that does not exist: that the reporter had written for The Atlantic. Four came from Perplexity, which named a specific, real, correctly cited article — by a different author entirely, verifiable on the Atlantic's own site. Two came from OpenAI, which cited nothing, here or anywhere else in the study.

Google AI Mode correctly rejected the claim all five times, and in doing so identified a similarly named Atlantic contributor — a different person the system had, correctly, not confused with the subject.

So in one corpus: a story the reporter broke, credited to two other journalists. A story someone else wrote, credited to the reporter. A citation can be entirely authentic — the right URL, the right publication, the right author — just not the author it was attached to. Attribution does not only erode. It moves.

Finding 4 — Where the connection forms, and where it doesn't

Fourteen of 25 responses said the two names belong to the same person. Ten said they could not tell. One said no, and described them as two distinct journalists who are sometimes confused with each other.

The split does not fall along the line I expected, and it is more interesting for that. Two platforms could not make the connection: one that answers from trained memory and one that answers by retrieving current pages. They failed for opposite reasons.

The memory platform had both halves of the career and would not join them. It correctly attributed the early outlets to one name and Computerworld to the other, then described two distinct journalists — and it was also the platform that correctly declined to guess at the INFOCON byline. It knew. It wouldn't assert.

The retrieval platform had the opposite problem. It found no record of the second name at all, on any run, and reported that nothing in its results connected the two — while citing, on all five runs, the very LinkedIn profile that lists the employer where those articles ran.

That profile is the single largest citation source in the study: 124 of 799 URLs, four times the next-most-cited source once you set aside the control question, which names its publication in the prompt. The page lists both employers under one name. It does not state the second byline. There is no shared string for a retrieval system to join on.

Publishing the history isn't enough. The succession fails at the identifier, not at the record.

And it does not fail only out on the open web. The successor archive that inherited this publication hosts two separate author pages for the same person, one under each surname, each with its own numeric ID and neither linking to the other. Both were reached during the study — different questions, different platforms. Nothing joined them, because there is nothing on either page to join.

Where the connection did form, it did not form on the archive. Three platforms affirmed it, and the citation trail on those answers runs through commercial people-search and contact-aggregator pages; one described its evidence as contact-detail continuity under both names, which is a public-records database, not a clip file. Seven responses asserted a combined-surname byline that appears in no publication and only in aggregator listings. The right answer, reached through the wrong evidence — and a scoring rubric that reads the text alone marks it correct.

Finding 5 — Wider distribution did not guarantee durable credit

Two of the reporter's stories carried further than the original outlet. One was picked up and hosted by a major broadcaster's site; it appeared five times in one platform's citations, describing the reporter's general coverage without attaching a byline to it. A separate set of pieces reached a national newspaper via wire syndication; across all 175 responses, that placement produced no signal at all — no citation, no mention, in any answer, on any platform.

That second result is worth stating precisely rather than interpreting: it is evidence of nothing, not evidence of failure. A null result is not the same as a negative one. What can be said is that broader distribution preserved a general description of the reporter's coverage in one case, and produced no retrievable trace in the other.

What the baseline says

The control question asked about a different Federal Computer Week reporter from the same years and the same beat. Zero of 25 responses identified her as a federal technology journalist. All 25 treated the name as ambiguous and asked for context.

That is not for want of a record. One platform cited her author archive at the same successor publication on all five runs, and still reported that the name referred to several different people.

That result cuts against the most dramatic reading of this study, which is why it is in the instrument. A working trade-press byline from the late 1990s does not reliably resolve on any of these platforms, for anyone. The absence documented here is not unique to one reporter. What is specific to this record is everything else: the fabricated attributions, the false confirmation, and the two names that will not join.

What this does not show

• That AI systems generally fail to credit journalists. This is one reporter's record, tested on one date, under one frozen instrument.
• That the subject should be well known to these systems. She is a private individual with a trade-press byline history, not a public figure, and low recall is an ordinary outcome. The baseline question was included to keep that possibility in view, and it did not rule it out.
• That the outlet with no retrievable signal handled the syndication worse than the outlet that produced five citations. A null result and a low result are not the same finding, and this study cannot distinguish “not indexed” from “indexed and simply never surfaced.”
• That the commercial aggregator sources behind some correct answers are reliable in general. This study observed that they exist in the citation trail for one name. It did not audit them.
• That absence is failure. The platform that declined to name an INFOCON reporter gave the only correct answer to that question.
• That no one reported INFOCON before March 19, 1998. That is a negative claim and no reprint can prove it. A search of contemporaneous 1998 security-press indexes turned up no earlier report naming the system, which is the strongest form available — and not the same thing as proof.

What I plan to do to improve

This is the part of the record that can actually be changed, and the reason to publish before acting on it — the correction comes first, same as every other AIWorthy study. The list follows directly from what the citations showed, not from what would be flattering to fix.

1. 1.Publish this study in full, including Finding 3, before making any changes. Done, September 8, 2026. This page carries machine-readable structured data naming both surnames as one author, pointing at both archives, and linking the underlying clips directly. The study's own argument is that no crawlable source states the connection — publishing it without supplying one would have been an odd omission. It is listed as an intervention, not a finding.
2. 2.Update the two LinkedIn entries covering this period so each one states the byline used at the time. That is the single missing string identified in Finding 4 — the profile records the employer, not the name printed on the work.
3. 3.Link directly to the clips referenced here and host permanent copies beside them, so the record does not depend on one publisher's archive continuing to exist. That includes a single page stating the INFOCON claim and its corroboration together: both headline forms, the outlet, the March 19 date, and the July 8 reprint. No such page exists right now — and the study showed the underlying articles are reachable without being retrieved.
4. 4.Point the existing ORCID researcher identifier at both names explicitly. ORCID is built to hold alternate name variants; a LinkedIn profile and a news archive are not.
5. 5.Rerun the identical, unmodified instrument in 90 days and publish whatever changes and whatever doesn't.

What I am deliberately not doing: chasing the aggregator records. They are carrying the name association that three platforms are using to get the connection right. The fix is a better source outcompeting them, not a thinner information environment.

The rerun

Next measurement: December 2026. Same seven questions, same five platforms, same scoring rules, same frozen instrument. The only variables expected to change are the ones listed above.

The first of those changes went live with this page on September 8. The December comparison is measured against that date rather than the collection date, and the change log is published so nobody has to take my word for the order things happened in.

Corrections and disputes

If we have described a figure or a platform's behavior incorrectly, tell us. We will check it against the source, publish a dated correction if we were wrong, and say so plainly. We correct errors of fact. We do not change findings because someone dislikes them.

AIWorthy is an independent research firm measuring how AI systems find, describe, credit and recommend companies, checked against the public record. We publish what we find, including where our own reference data was wrong. We do not grade a remedy we have been paid to deliver, we do not audit a party we have been paid by, and every commercial relationship is disclosed.

Footer language updated 14 September 2026 to reflect current service scope. No findings, figures, methods or artifacts were changed.

hharreld@aiworthy.ai · aiworthy.ai