Access Isn’t Trust
AI aggregation without accountability corrupts knowledge
Access to words, thoughts and information are no longer a bottleneck. Access to trusted resources are. In this new era of conversational bibliometrics and AI-mediated discovery the responsibility for these research resources now sits squarely with data and infrastructure providers.
Knowledge is shaped by infrastructure and data
In reading friend and colleague Dr Helene Draux’s piece on conversational bibliometrics, I began to think more about how data fits into the conversation.
Her point is that the new wave of conversational AI interfaces removes technical barriers. Researchers no longer need to form a hypothesis, then conduct a literature review, thinking about the restraints of the query needed to field our question. We don’t need the technical expertise of working with APIs – knowing the data schema, understanding the data limitations, figuring out if it’s a ‘them problem’ or a ‘me problem’. BUT asking those questions comes with trusting the infrastructure the system is built upon. We have ease of information but not necessarily understanding. And our insights could be built on sand.
Her point is that methodological responsibility must now be embedded into the system itself for sound research systems. In conversational bibliometrics, trustworthy infrastructure is vital. That is half of the equation.
Trusted data – explicit data governance, provenance tracking, and embedded trust-weighting mechanisms – complement conversational systems’ infrastructure thus reducing the risk of amplifying low-integrity research at scale.
Hélène makes a great point for verifiable infrastructure built with sound practices. I make the point that you also need robust, trustworthy data. And that responsibility includes data quality, provenance, and trust signals, in addition to schema design. (Just to be very academic here – ‘data’ can mean different things including in the numeric sense as well as articles or journals as data points.)
Structural Tension - Aggregating Research
Let’s take a step back and understand the change in research publications. Briefly, the scholarly record moved from journals to aggregation systems. This includes in the aggregated research intelligence platforms (Dimensions, Web of Science, etc), which depend on choices of what data to include and exclude. This moved knowledge discovery to individuals, bypassing libraries for curation.
Web of Science, as one example, bases the aggregation at the journal level and has opted to curate and publicize which journals make their integrity cuts. This is a helpful step – if we still look at trust at the journal level. The thing is, some journals can be very bad – all the way around. Predatory journals. But there is the nuance of some journals being bad some of the time. Then there are some articles that pollute journals, but the journal itself remains strong.
That actually lends strength to the Dimensions approach of collecting scholarship at the article level. It also allows some smaller journals and their articles that may not make the cut for Web of Science into a system for wider consumption – vital for robust, global science. Unfortunately, this ingestion method is also principled on minimizing curation oversight, which allows a lot of junk into the system (but also excludes predatory journals’ ‘publications’).
No approach is perfect. And all of the current tools still don’t address the challenge that in conversational bibliometrics, we need to be able to weight the trust in an article. Not the science, but before the science – the trust in the person authoring the work and the transparency of the research environment along with the journal properties.
This is critical in the age of regurgitation of information into AI summarization tools.
The New Problem: Problematic Data & AI Output
AI removes friction. It removes the need for query expertise. It removes bibliometric literacy from the consumption layer. And in doing so, it multiplies exposure to unweighted and potentially problematic information.
AI brings all data together using systems described above and summarizes topics into seemingly believable narratives. So what would a system look like if you had sound infrastructure but unreliable data? A few examples:
Problematic Data
Data ingest problems. As noted above, Dimensions does not curate by journal (same with OpenAlex) but at the article level. Yet what happens when a couple of problematic ‘publishers’ appear in the databases? All of the articles are ingested. Darcy and Roy Press (covered by science journalist Jack Ryan with more detail) and Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP) are two examples. These problematic places came in from upstream systems - getting through their checks then ingested into systems that trust their processes. More on these in the coming weeks.
AI Data Out: New places are popping up all over offering AI tools for ‘trusted research’. Sourcely, AnswerThis, in addition to Google Scholar Labs. So I decided to do a quick check with their results based on a question.
My prompt: What are the top articles on Alzheimer’s research coming out of Bangladesh? Maybe not the best-worded prompt (what does ‘top’ really mean?) but I asked this question for a reason. I know a lot about one problematic network where the lead person hails from this country. I also know there are a lot of good researchers in Bangladesh – yet the researchers and research may be overlooked because of the country’s economic status.
The two sources, Sourcely and Google Scholar, returned at least one paper that is not retracted but part of a known authorship-for-sale network. AnswerThis didn’t really offer much on the trial as I could only see a few articles without subscribing to and paying for ‘Pro’.
Problematic data coupled with Biases
Besides one place returning a dodgy paper, it also put its biases directly in the summary.
What I wasn’t expecting was this response. Before the tool returned the suggested research papers, this summary appeared: “Undefined Bangladesh has limited prominent research output specifically focused on Alzheimer’s disease compared to major research hubs in developed nations. While there are some neuroscience and medical research initiatives in Bangladesh, comprehensive… (cuts off there unless I paid for pro, which I did not.)
This comment is problematic on so many levels. However, at least the biases are visible. This is a case of visibly showing the biased algorithms regardless of the available data.
May I remind everyone that we need research throughout the world? Research is global. Legitimate science emerges from every economic context. And yet, conversational systems may embed geopolitical priors before ever presenting evidence.
While we’re somewhat on the topic, people and organizations who game the system through papermills, data manipulation, citation cartels, etc. are all over the world. And the gaming hurts the people trying to conduct legitimate research.
Final thoughts: Is monitoring the same as policing?
Conversational bibliometrics will and should shift expertise upward, into infrastructure and data governance layers. If users no longer need to understand schemas, query logic, or database limitations, then those responsibilities do not disappear. They consolidate.
So who is responsible for trusted data?
Not the end user. Not the researcher typing a prompt into an AI interface. The responsibility lies with the many publishers, institutions, funders. AND trusted data must be embedded within the systems that ingest, structure, classify, weight, and summarize research outputs.
This brings us to the tension: is monitoring the same as policing?
There remains discomfort in parts of the scholarly community around filtering and flagging research outputs. Monitoring can feel like gatekeeping. But there is a critical distinction between suppressing content and surfacing contextual risk signals.
Open data is not the same as trustworthy data.
Access is not the same as comprehension.
Monitoring and prevention are not censorship.
Prevention is governance.
Monitoring is infrastructure.
If conversational systems are going to synthesize, summarize, and narrate the scholarly record at scale, then trust cannot be assumed. It must be baked in. Because aggregation is not the same as accountability.
If we fail to build article-level trust-weighting, transparent provenance, and explicit governance into conversational bibliometrics, we risk building beautifully engineered systems on contaminated sand.
And once AI systems regurgitate that tainted sand at scale, it becomes very difficult to contain.
 | A guest post by
|
"Access to words is no longer a bottleneck. Access to trusted resources are." — this is the problem statement for a layer of infrastructure that doesn't exist yet.
I've been building it. A hash-chained event graph where every claim links to its source, every source has a verifiable track record, and trust is weighted — not binary. A researcher whose claims consistently survive challenges accumulates credibility on the graph. One whose work is retracted or debunked loses it. Not as a rating someone assigns — as a pattern that emerges from the chain.
Your point about needing trust at the article level rather than the journal level is exactly right, and it maps to a broader architectural principle: trust should be derived from behaviour, not from membership. Being published in a good journal is membership. Having your claims survive independent verification is behaviour. The first can be gamed. The second can't.
The Knowledge Graph layer of the architecture I'm building does exactly this — claim provenance, challenge events, source reputation derived from history. 38 posts on the full architecture at mattsearles2.substack.com.