Texas Bar Journal • March 2025
Large Language Models
A look at their role in e-discovery.
Written by Chase Hardy and Matthew J. Hill
Attorneys have long been known as “merchants of words”—we trade in language, build arguments around text, and parse meaning from documents. From drafting complaints and motions to reviewing thousands of pages of discovery, our work demands both precision and efficiency in the interpretation of words. Our profession now faces a pivotal moment: Should we, as counsel, incorporate artificial intelligence to assist our discovery processes—and if so, how quickly must we adopt it?
FROM MANUAL REVIEW TO TECH-ASSISTED REVIEW
In the early days of discovery, reviewing documents was an entirely
manual process. Attorneys would spend countless hours sifting through
boxes filled with physical documents— contracts, emails, letters,
billing statements, and more. The goal was (and still is) to identify
information that might be relevant to the issues in a case. However,
the sheer volume of material that a single attorney could reasonably
read in a day, week, or month placed obvious constraints on how
thorough and cost-effective this approach could be.
When more voluminous requests arose, some law firms attempted to streamline manual review by offshoring large-scale document review projects. Yet this solution was hardly ideal. Documents were still subject to human error, inconsistencies in review standards, and potential confidentiality concerns—not to mention the enormous costs that inevitably arose from paying a specialized labor force.
Enter technology-assisted review (TAR). Over the past couple of decades, TAR has introduced more sophisticated ways to sift through a mountain of electronically stored information.1 Early TAR tools relied heavily on keyword search. Counsel would feed in a list of key terms—names, product references, or significant events—and the software would retrieve documents containing those terms. While these tools were far more efficient than manual review alone, attorneys often noticed that keyword-only methods would bring back false positives (due to overbroad search terms) and still miss documents containing synonyms or differently phrased content. The technology is also highly niche and commonly inaccessible to the small practitioner.2
With the rise of user-friendly large language models (LLMs) such as ChatGPT, Claude, Gemini, Grok, and other general consumer-facing AI models, the discovery landscape has shifted yet again.
EMERGENCE OF LARGE LANGUAGE MODELS
Large language models represent the next frontier in discovery
review. They are powered by billions of “parameters,”
which are essentially the numerical values the model learns and adjusts
during training to detect patterns in language.3,4 Rather
than searching for exact keywords, they glean meaning from context,
synonyms, and phrasing. This means that an attorney can ask an LLM to
locate all documents discussing a certain topic—even if the topic is
never spelled out by the same words. For instance, a user might prompt:
“Retrieve all emails that mention dissatisfaction with a specific
product launch,” and the model might find references to
“problems,” “issues,” “glitches,”
or “complaints” in addition to the word
“dissatisfaction.” A search through social media history for
“vehicle” may return posts referencing “cars,”
“trucks,” or “motorcycles.”5
ADVANTAGES
OF
AI
FOR
E-DISCOVERY
Finding What You Don’t KnowYou’re Missing
One of the most promising advantages of AI in discovery is its
capacity to find things you never knew to look for. With keyword
searches, you are limited to what you can guess is important. If you
do not use synonyms for “contract,” the system will not
catch references to “agreement” or “memorandum of
understanding.” Large language models do not rely on rigid syntax
alone; they interpret semantics, focusing on meaning rather than just
words. It’s possible to type to an LLM in the same way an
attorney would type to an associate. This dramatically reduces the risk
of missing key documents simply because keywords were phrased
differently.
Sentiment Detection
LLMs can detect sentiment—whether a speaker is angry, frustrated,
pleased, or suspicious. This can be pivotal in high-stakes litigation.
For example, a cryptic email from a corporate executive might hint at
wrongdoing, but unless you read every word from every email you might
miss it. An AI-driven approach can detect negative or secretive
undertones that suggest a deeper problem. By highlighting these
nuances, lawyers can spot red flags faster, leading to more informed
case strategies. Moreover, AI review can quickly discover communication
patterns that keyword searches, and even human review, may miss.
Time Savings and Cost Efficiency
Time is money—for both the attorney and the client. A thorough
manual review of large volumes of emails and documents could take months
or even years, rack up enormous billable hours, and burn out staff.
AI speeds this process exponentially, reviewing tens (or hundreds) of
thousands of pages in a fraction of the time. The result is a win-win:
attorneys can handle more cases and clients save money on the overall
cost of discovery.
Repeatability and Consistency
Another boon of AI-assisted document review is repeatability. Large
language models retain the same level of diligence every time you rerun
a search or train a new model. Humans, on the other hand, can miss
details on a second or third pass due to fatigue or oversight. AI
ensures consistent results across a standard model.
CONTEXT WINDOW: THE LIMITING FACTOR
A defining feature of large language models is the “context
window,” or the amount of information—beyond the model’s
general training—that you want it to process at once. Early large
language models could only analyze a few thousand tokens (a token is
typically around a few characters). Newer consumer-grade large models
can handle hundreds of thousands of tokens, corresponding to hundreds
of thousands of words—enough for entire pleading sets, demand letters,
discovery responses, emails, contracts, medical records, and more to be
parsed and processed simultaneously.
The impact of increasing context windows is remarkable. Today, large language models can draft bespoke discovery requests and respond in kind, and they are rapidly becoming multimodal—capable of processing not just text but also images, audio, and video. As a result, the only real limitation for lawyers is the size of the LLM’s context window.
The CEO of OpenAI (maker of ChatGPT) suggests that, in the near future, context windows will expand to natively accommodate many billions of words.6 This expansion means attorneys might eventually feed every bit of information even remotely related to a case—hundreds of thousands or even millions of pages—into a LLM at one time.
MOVING TOWARD A DUTY TO USE AI
Raising the Standard of Practice
As these models become more sophisticated and ubiquitous, the legal
industry must confront a pressing question: Does failure to use AI tools
in document review risk a violation of the attorney’s duty of
competence? The current proposed ethics opinion says no.7 But
what does the future hold? Many practitioners and ethics boards are
beginning to assert that ignoring AI might fall below the standard of
practice—especially in complex litigation where vast amounts of data
are at stake.8,9 If cost-effective AI solutions can locate
relevant material that might otherwise remain hidden, an attorney who
proceeds solely with manual or outdated review strategies might open
themselves up to malpractice claims.
Ethical and Practical Considerations
Of course, adopting AI does not absolve attorneys of the
responsibility to ensure confidentiality or compliance with privilege
rules. Lawyers remain the gatekeepers for what is produced. Ethical
obligations such as verifying the accuracy of the documents identified
by AI, maintaining client confidentiality, and safeguarding privileged
material remain paramount. If an AI tool is used, it is incumbent on
attorneys to verify the reliability and security of the platform. Many
large language models allow for secure cloud deployment, mitigating
concerns about data breaches or accidental sharing of confidential
information.10 However, due diligence in selecting an AI
provider is crucial.
Access to Justice
When used responsibly, AI can also enhance access to justice by
reducing legal fees. Small firms and solo practitioners who previously
found large-scale e-discovery cost- prohibitive can now leverage
AI-driven services. By lowering financial barriers, more clients can
afford to pursue legitimate claims or defenses. In this way, AI may not
only redefine best practices but also reshape how we serve a broader
population in need of legal representation.
CONCLUSION
In a profession defined by words, large language models represent a
natural evolution of our craft. Rather than an assault on our
traditional role, AI is becoming an indispensable ally, freeing us to
do what we do best: develop legal strategies, counsel clients, and
focus on the higher-level skills that define our calling.
For those who remain skeptical, consider this: as context windows stretch to encompass billions of words, as AI systems learn to cross-reference details at a granular level, the concept of what is “discoverable” will change dramatically. The time spent in data review will shift to time spent in strategic thinking. Clients will be better served, and attorneys will be more effective. Indeed, in the not-too-distant future, AI-assisted discovery will likely become a baseline expectation—a standard tool that every attorney is expected to command. To stand still is to fall behind. The future is now, and it speaks in ones and zeros—but it speaks our language all the same.
