It’s no secret that Artificial Intelligence (AI) is already beginning to revolutionize the human experience, with profound and far-reaching effects that are not yet fully understood, and in some cases probably have not yet manifested. That includes the field of historical research, where TRG is already seeing AI’s impact. We are also starting to receive questions from clients and prospective clients regarding our use of AI and how it can streamline our work.
TRG is a 21st century company that enthusiastically embraces new technology. From the early days when laptop computers, flash drives, portable scanners, and ready access to OCR felt like exciting workplace innovations, our team has spent more than 25 years making full use of an ever-expanding universe of digital imaging tools, document storage systems, and online resources. We love the smell of old books and still relish the sight of old-school card catalogs, but we pride ourselves on being tech-savvy researchers who take advantage of all that the modern age has to offer our profession.
It may come as a surprise, then, that although we have readily incorporated AI into our researcher’s tool set, it does not play a major role in our day-to-day business. That is due to several critical limitations in AI’s current capabilities that in turn limit its usefulness in our line of work. This blog will address the most important of those AI limitations as they currently stand, key among them being an absence of historical information available to AI programs.
When it comes to history, the vast majority of records are still analog.
When our clients turn to us, they are seeking detailed historical research to address specific and sometimes complicated questions, not the basic types of historical summary that can be found via a simple Google search or by consulting published overviews. Indeed, most of our clients have already done that preliminary work, and often more, when they hire us. That means that most of our work involves a deeper dive into obscure and otherwise banal material that has not yet been digitized or analyzed thoroughly in digitized secondary sources, and therefore is currently completely outside of AI’s access because it cannot be included in the datasets that underpin AI models.
The vast amount of digitized information that is currently available can give the false sense that most material has been digitized, or soon will be, but nothing could be farther from the truth. For example, in 2025 the U.S. National Archives and Records Administration (“NARA”) reported having digitized 444,580,681 pages of records. That is obviously a massive volume of information, yet it likely accounts for less than 3.3% of NARA’s textual holdings, which include over 13.5 billion pages of paper records, not to mention its more than 10 million maps and well over 41 million photographs.
Digitization is costly and time-consuming, so when choosing what material to digitize, archives—whose budgets tend to be limited—often prioritize records that are popularly requested or too fragile for physical handling. In our experience, the vast majority of historical records that are of value to our clients—such as routine interagency and company correspondence, site-specific reports, or documents related to specific purchases of raw materials or obscure equipment subcomponents—are not those considered to be in high public demand (unlike, say, the papers of the nation’s Founders or records valuable in genealogy).
Therefore, not only is a massive trove of records that has been critical to our clients not currently digitized, but it is not likely to be digitized anytime soon, and is likely to remain outside of AI’s access for the foreseeable future. While NARA has committed to increasing its volume of digitized records, full digitization is a long way off, especially in the wake of recent budget and staffing cuts. While a large portion of our research is at NARA and other federal repositories, practical limits on digitization are nearly universal, which is why we routinely conduct research at repositories nationwide.
AI’s underlying datasets are only as good as their source material.
Digitized records that make their way into the various AI datasets can also create problems. They, too, can contain errors or omissions, sometimes exacerbated by the fact that digitized public domain materials that began life as analog (i.e., printed) publications whose copyright has expired are more likely to be based on outdated scholarship that is inaccurate or incomplete. Some more recent scholarly publications, although available in digital form, may still be beyond AI’s access due to copyright protections and private access restrictions. The issue of intellectual property owners being adequately compensated for their work being fed into AI models is, of course, contentious and ongoing.
Gaps and errors.
AI’s lack of access to the full historical record does more than just create gaps in its output; it also leads to a high probability of outright errors. Analysis of any sequence of events where one of the events is unknown can lead to serious misunderstandings of history, and AI has demonstrated an inability to identify the gaps in its own knowledge. Access to the complete historical record is also the best way to combat the presence or predominance of erroneous information, an issue that we already dealt with routinely before AI entered the mainstream. There are, of course, errors and omissions in historical documentation, and the best way to identify them is by analyzing as much of the extant historical record as possible.
But AI errors are not solely due to data gaps. They are sometimes due to flaws in (or features of) the AI programs themselves, such as chatbots that can synthesize information but are also essentially predictive text generators. These can potentially result in misinterpretation of the data on hand (including the models’ inability to separate fact from belief), “AI hallucination,” and the subsequent tendency of AI programs to be “confidently wrong”). They can also be caused by intentional human manipulation. Considering all of this, it is easy to see why we tread carefully.
The bad AI historical summaries that we routinely encounter range from vague and unhelpful to laughably inaccurate sources of outright misinformation. For example, while conducting preliminary research for a work plan in 2025, a Google search produced an “AI Overview” claiming that a company of interest’s “correspondence, financial records and legal documents” were held within a university’s special collections library. We leave no stone unturned in our research and therefore felt obligated to follow that lead and make inquiries to the university, only to find that it held no such records and that the collection did not exist. In that case, the misinformation caused hardly any delay to our research, but the cost of chasing invalid leads can add up over the course of a project. The precise reason that Google’s AI Overview provided this false information is unknown to us, but in our experience this type of error is not an isolated occurrence, and in fact each Google AI Overview contains the disclaimer “AI responses may include mistakes.”
There is still a place for AI in historical research.
So how do we use AI? Like any other tool, AI can be useful, but for us it is another means of finding original documentation, not an end product. If the underlying dataset is sufficient, AI programs are capable of quickly summarizing information in a helpful way, saving time. That is particularly true when it comes to summarizing historical context. If we are looking for a quick summary of a historical event that may have had an impact on a site or company we are researching, a paragraph or two from generative AI may be just the level of detail we need, if fed by digitally available content of sufficient quantity and quality to be at least broadly accurate. AI summaries that include citations often lead us to other records of interest. However, at TRG, AI is never the last word.
At TRG, trust is paramount.
Our clients trust us to provide 100% accurate documentation and fact-based analysis, which often finds its way into the courtroom or other legal settings. Misinformation, including “phantom citations,” is already beginning to plague the legal world, with potentially disastrous consequences. Therefore, TRG’s research, analysis, and end work products will always be steered and quality-controlled by our expert staff, and the records we collect will always be fully documented to their point of origin. We will always seek ways to meet our clients’ needs in the most efficient and cost-effective manner—a trait for which we are known—but TRG will never take shortcuts at the risk of inaccuracy.
Please note: no generative AI was used in writing this blog.
