The Power of Nothing in AI and Law

The Danger of Everything, Value in Void + Extrinsic Evidence, Authentication Law

I always thought it was too much when people, most often academics, made overlong titles for white papers and articles.

Yet here we are today.

This is necessary, for clarity, to be stated at the outset since its a novel think and to show you where we are going (and where I recommend you go in thinking about it) :

The Power of Nothing in AI and Law,
And the Danger of Everything: AKA The Value of the Void

How Null Answers in Law and AI Reveal the Path Forward,
And How “Right Answers” Can Mislead

Extrinsic Evidence + Authentication

In both the Law and AI, the concept of a “null” or empty answer (no value) is most often misunderstood.

I think far far more in AI than Law actually. And why Law is here to help AI.

Many might see a lack of information (a zero) as a failure/fault. On a level they are right.

Some people get downright angry, “Sucks! Didn’t work.”

But in another reality, a null value can be a critical indicator of where to focus next. Null has value. Sometimes, and in my experience more often than not, it has GREAT VALUE — more than other “actual answers.”

And this is true very much in Law — and also in AI as well. NOTHING CAN BE EVERYTHING.

This “space in between”—the absence of data or a clear and final answer—holds immense value because it points directly to the missing component or the next step that requires attention and/or conceals (but POINTS AT) something needing development.

Here are three examples :

1. That thing your client never told you, until during a brief break in a deposition. The “null” value (even when unknown) HAD A MATERIAL IMPACT on the direction of the case. Now aware, you must “adjust” and do — to get closer to win. The value of the “null” was and is everything.
   

2. A vaguery in a contract you are reviewing. Something unclear. It is a “null” a “gap” a “miss.” It could just be a mistake. More frequently in my own experience it means (a) someone doesn’t know (b) someone doesn’t want me to know (‘c) the value itself is unknown (and I can provide it / define it). One could make an argument that in contract review these are the “Easter Eggs” - the things of great value. Nothing is everything.

In AI, your typical user will cry (often like a baby without its bottle) in the face of one of these nulls/vagueries. “Didn’t work.” “Hallucinated.” “This model sucks.”

AND YET IN AI AS LAW THE OPPOSITE COULD BE TRUE - NOTHING COULD BE EVERYTHING

3. =

A Story from X

Recently I did an AI CHRONOLOGY Crunch on the V versus Apple and OAI case. Took just the complaint

T/he following day, xAI Grok released a lawbot on the case where users could “chat” with the data.

Had a dialogue with a fellow user on X about both of these.

He said mine “did not have the answer”
and that his (“gave me the answer”).

Except that his answer was wrong and correct and mine was right and correct.

He began by saying my AI Digest “doesn’t seem to be in a flow”

[Mine had listed (just) ‘November’ for one event and November 17 for another. Viewer was bothered by the “no flow” (assume this means, not complete or the month/date out of synch)/]

My reply (answer was correct as doc only provided the month not the specific date — and so the AI output answered, correctly, that the value was in fact “November”)

“The lack of (precise) date is data”

User “Ashutosh” replied that my explanation “that is clear!” (insinuating the prior AI output had not been) and that the other tool “arrived at the answer”.

ITS VERY INTERESTING BECAUSE THE GROK LAWBOT WAS TO BE BASED ON THE COMPLAINT ALSO /// MINE MOST CERTAINLY WAS ONLY AND EXCLUSIVELY THE COMPLAINT // IF GROK CAME WITH A DATE CERTAIN WHICH MINE DID NOT

HOW DID THIS HAPPEN?

THIS MEANS IT RETRIEVED IT FROM ANOTHER SOURCE - NOT THE DOCUMENT

IN LAW WE CALL THIS “EXTRINSIC EVIDENCE” AND IS GENERALLY DISFAVORED AND INADMISSABLE

And my reply (text continues here in body of article italicized)

“You’ve got me hustling so thank you

just did a key search of the pdf and guess guess what?

date of nov 21 does not appear in the document.

repeat, the date of nov 21 does not appear in the document.

EXTRINSIC EVIDENCE in the law, we have something called “extrinsic evidence.” what this means is something outside of the record used to prove something inside of the record. we know that the db (database) used in the edubot, xai’s first lawbot,

contains extrinsic evidence

since the information [the date] does not appear in that document.

likely it checked with the grok db (database).

the db for ours was the doc and only the doc.

this is not bad nor good again it’s just information as we craft best tools for best use cases - here law.

the document itself (at p 11, 48) says only “a few days later”

-/ which is imprecise

1 the analysis i provided was correct as to the document which did not specify a date certain (only the month of November)

2 an attorney perhaps in a proceeding, were he/she to cite the date you provided, would be submitting extrinsic evidence…not part of the record… (= IMPROPER)

3 in working with llms, we deal in probabilities. (AND HUGE OFTEN NON-LEGAL DATASETS)

4 in working in law, probabilities are perilous and fraught with error.

5 you will note that ours (UNLIKE THEIRS) provides (HYPERLINKED) citations to the source document so reviewer can go look — and here would have seen “a few days later” to answer the question the analysis itself posed.

6 in our tool these are direct links. so click citations and it brings you not just to the doc but to the line where the information comes from (and we highlight it).

7 sources of information (provenance) are essential in law. judges and opposing counsel will always say “where do you get that” (because they too have reviewed the same documents in detail).

8 major llms and almost all ai for law do not provide sourcing for outputs (much less direct linking) as our tools do.” (AND HAVE ALL SORTS OF “JUNK DATA” (LIKE REDDIT) IN THEM)

***

Even though my collocutor did not have any reply, I thank him for sparking with his push-back some deep break through thinking and confirmatory insights on building modalities for AI in Law.

Nothing can indeed be everything in Law and AI. And everything can be nothing.

Here — the Complaint was imprecise. The LLM provided data “that it should not have.”

=== ours provided the null / the gap which is important “signal” to the practitioner to drill down in that area. Without the null being flagged, practitioner advances misled and unaware of the misleading!

* * *

The pursuit of “everything” in AI, particularly through giant large language models (LLMs), with as much and any data as computationally possible, carries great DANGERS in Law — especially when it leads to the importation of extraneous or fabricated information. (“extrinsic evidence”)

Again, The Power of Nothing in Law

In Law, probabilities are perilous, and certainties are paramount.

In legal practice, the absence of information is not a dead end but a signpost.

As one perspective highlights, “Lack of data is data.”

This means that when a legal query or analysis yields no direct answer, it doesn’t signify failure; rather, it signals a gap that demands further investigation.

Smart lawyers understand that these voids are opportunities to drill down into specific areas, uncovering valuable answers that might otherwise remain hidden. (And if ignored due to reliance on an “easy answer” the lawyer most always ends up paying an expensive price later.

WE HAVE TO KNOW WHAT IS UNKNOWN. ESPECIALLY THOSE THINGS WHICH ARE KNOWN TO BE UNKNOWN!

Contrasted with AI

Silence or “nulls” in AI are treated as abject failures of the AI. Like “dead air” on a radio station - catastrophic.

Yet this bias toward expecting AI to provide exhaustive answers (or just “any answer” / never a non-answer) can be dangerous, as it may lead to overlooking the importance of what is unknown — or the infilling of an answer from a source which is not good.

In this case “everything” (it told me all this stuff) is really “nothing” at the same time.

Extrinsic Evidence

And this data based on junk sources has an additional problem beyond being wrong.

As was seen in the ‘your AI v my AI’ example above,

This evidence lies outside the official record, or the record being considered.

AND MOREOVER, YOU DON’T EVEN KNOW IT. WHICH IS A MUCH MORE DANGEROUS FORM - WHEN YOU “THINK” IT IS PART OF THE RECORD BUT IS NOT!

The null answer, in Law and AI, acts as a reminder to stay within the “four corners” of the document (or within “the record” of the case or matter) unless there is a compelling reason to look elsewhere — and PERMISSION GAINED. It enforces accuracy in sources, and clarity in the agreement that we are all considering the one same record.

Extrinsic evidence can be cheating. Creating a “shadow” record/reality — one that neither opposing counsel nor judge are playing in! It is unfair advantage which destroys the integrity of the proceeding.

Treating Null Answers in AI for Law

The same principle should apply to AI, particularly in professional applications where accuracy and reliability are non-negotiable.

In AI systems, a null answer is not a defect but a success in identifying what is not known.

[This is a stark contrast to recreational AI, (90% of AI?) which often prioritizes providing an answer—any answer—over admitting uncertainty. This is “AI for fun” —where the stakes are nothing.]

Professional AI, however, must recognize that the absence of data is itself data. And IMPORTANT data.

It succeeds by highlighting the gaps, prompting users to take the next step, whether that involves gathering more data, refining the query, or exploring alternative approaches.

A Word on Authentication

If our large LLM data sources are unknown, unattributed, and uncited — and most are, then the Extrinsic Evidence (and frankly all the outputs) pose another problem:

AUTHENTICATION

In Law we need to know that evidence used is genuine. LLMs offer no such assurance.

In fact what they do, in that we know they provide a “jumble of information” from various unknown sources, what they do is produce outputs

INCAPABLE OF AUTHENTICATION (!)

Even if there are “good sources” underlining that output, not only do you not know, but technically you cannot know, its not possible. (without citations)

Judges and opposing counsel frequently suddenly maybe daily demand, “Where do you get that?” Remember they are in as deep into the record as you are. And you had better have an answer.

With the extrinsic evidence + authentication problems endemic to LLMS = you don’t and you can’t.

AI users (for Law applications) must be able to trace the origin of information.

More Trouble with LLMs

Want some examples of unattributed EXTRINSIC EVIDENCE which is not AUTHENTICATED causing big problems in LAW for lawyers due to the “black box” junk data of big box LLMs?

Here are three :

1. The Case of Ko v. Li (Ontario Superior Court, 2025)
In early 2025, lawyer Jisuh Lee submitted a legal factum in the case of Ko v. Li that contained citations to two non-existent cases. These citations were later discovered by Judge Fred Myers, who imposed sanctions on Lee. The error could only have been generated by an LLM trained on flawed or incomplete data, likely drawing on internet-sourced legal summaries or mis-indexed case law that led to the creation of entirely fictional citations.

2. The Morgan & Morgan Sanctions (U.S. District Court for the District of Wyoming, 2025)
In February 2025, lawyers from Morgan & Morgan and the Goody Law Group were sanctioned for submitting a brief with fabricated case citations. The brief included descriptions of legal standards that were “peculiar” and unsupported by any existing precedent. The generation of these fake citations likely stemmed from an LLM’s reliance on bad data or a combination of flawed datasets, incorporating mislabeled or incorrectly summarized case law.

3. The Ellis George LLP and K&L Gates LLP Fiasco (U.S. District Court for the Central District of California, 2025)
Attorneys from Ellis George LLP and K&L Gates LLP submitted a brief containing numerous hallucinated citations, discovered during a hearing. The LLM likely accessed a combination of real case law and erroneous internet-sourced content, creating a hybrid of accurate and fabricated information. The lawyers’ failure to recognize the inconsistencies suggests a reliance on AI output without critical evaluation, exacerbated by the LLM’s flawed training data.

GUYS THIS IS K&L GATES - “NO SLOUCH” AS WE SAY!

These examples underscore a these critical

DESIGN DEFECTS

OF USE OF LLMS AND “LLP WRAPPERS” FOR LAW

LLMs are trained on vast corpora that include legal texts, but these datasets are not immune to errors, biases, or outdated information. “Wrappers” are programs built atop an LLM which make use of the same LLM database in ways they themselves program/

A significant portion of this training data comes from non-legal sources, such as Reddit, YouTube, and other internet content, which may contain inaccuracies, opinions, or fabrications.

For instance, Reddit threads often include anecdotal evidence or speculative legal advice, while YouTube videos may feature misleading or simplified explanations of legal concepts.

LAY PERSONS OPINING LAW WILL BE CLASSES LAW DATA BY THE LLM!

When LLMs incorporate this extraneous data, they inevitably generate outputs that are not only irrelevant but also false and potentially harmful in a legal context.

The danger of “everything” in AI is real lies in its tendency to overpromise and OVERdeliver (without context, source, or specificity……….much less provenance or citation.)

Conclusion: Embracing the Null - Nothing is Everything

The “space in between”—the null answer, the void, the absence of data—is not a barrier but a bridge.

In both law and AI, it points to the next step, the missing component, the area that requires attention.

By embracing this space, professionals can transform uncertainty into clarity, gaps into insights, and nothing into everything.

As the adage goes, “Nothing is everything,” and in the contexts of law and AI, this truth is more relevant than ever. The null answer is not an end but a beginning, a call to action that drives us toward deeper understanding and more precise outcomes.

HOW DO WE DO IT? IT IS NOT ROCKET SCIENCE - ITS SOUND LAW AND LEGAL PRINCIPLES EXPRESSED WITH CODE

1. Legal datasets. Don’t use “junk data.”

2. Labeling outputs and citations for data.

3. “Optional and User-selected-only” Internet search.

4. “Optional and User-selected-only” LLM usage/reference for research.

5. Direct links to data source/s, and to specific section/s of source so lawyer can “QA the AI — concordant with his or her ethical duty to do so. We must check our AI.

We do all these at Jurisphere. Unlike others. And it makes a great difference. And it will change the world, or better, create the world of AI for Law. It already is.

Truly Legal AI.

CODA

The dozen+ bespoke lawbots I have yet built for Project gist (and those we are yet building) are are all this.

1. SLM not LLM - Just the Law data we need nothing more.

2. LAW - No non-law sources.

3. SPECIFIC - You don’t need one area of law in another. Specific wins.

4. GIVES NULLS WILLINGLY. These bespoke lawbots know how to say No. “I’m sorry that is not in my record set.” We treat this as a “W” each time it happens. It means probably the question/prompt was out of scope — and another user was not misled.

5. GIST BESPOKE LAWBOTS CANNOT TELL YOU WHO BUILT THE EIFFEL TOWER. NONE OF THEM. NOT A SINGLE ONE. WE CALL THIS THE “EIFFEL TOWER TEST.” IF ONE OF MY LAWBOTS GIVES ME THIS ANSWER, I KNOW SOMETHING IS “OFF.”

This is the brown M & Ms in the bowl for Van Halen. IYKYK