I presented at an industry conference in 2019. From the stage, I asked a room of 400 litigation support and eDiscovery professionals:
“How many of you are using predictive coding as a routine workflow?”
Four hands went up.
Four. Out of 400.
This was 10 years after predictive coding became widely available, yet the adoption rate was abysmal. If you had asked the other 396 people why they weren’t using it, you would have heard the usual (and very confident) objections:
- “But how do we defend it?”
- “If we don’t know exactly how it makes decisions, can we really trust it?”
- “Courts won’t accept a black box!”
- “I need to see case law permitting its use.”
- “Opposing counsel won’t allow it.”
- “We don’t want to exchange seed sets.”
And now, we inhabit a world even more saturated with Generative AI. And we are hearing the exact. Same. Arguments.
With two new catch-all additions:
- “But it hallucinates.”
- “We don’t know where the data is going.”
Concerns so grave that some liability carriers have instructed law firms not to use GenAI—because, apparently, young associates have never misused work product or broken confidentiality on the internet in the last 20 years…
But really, we’ve seen these challenges before. And adoption does grow—albeit slowly.
Here’s the thing about GenAI: If you’re comfortable defending keyword searches, TAR models, and search term reports, you should be waymore comfortable with an AI-augmented decision on document relevance or privilege.
It all comes down to defensibility and validation—two vastly different but equally important concepts.
Defensibility = Can you justify your process? Validation = Can you prove it works?
Unfortunately, you CAN have one without the other.
Validated but Not Defensible
If you’ve made it this far into this article, you’ve likely been involved in an eDiscovery project. Which means we can talk about the fine art of modifying keyword searches until the results ‘look about right.’
You know the drill:
- Start with a set of search terms (of vague origins…)
- Realize the results are too broad.
- Start tweaking.
- Throw in a proximity operator.
- Narrow the scope.
- Remove a word that’s pulling in too much junk.
- Keep adjusting until the hit count seems reasonable.
- Declare victory and pretend this was science all along.
This is NOT defensible—and would be extremely uncomfortable to explain at a Motion to Compel hearing.
Defensible but Not Validated
You’ve got a process. A beautiful, documented, multi-step QC workflow.
It looks airtight. Bulletproof. Something you’d proudly walk into court and defend.
And then… someone actually checks the results.
- Turns out the review team wasn’t properly trained on key issues.
- (Turns out the associate conducting the review team training wasn’t properly trained on the issues…)
- Half the documents were misclassified.
- WAY too many documents were tagged as privileged.
But hey, at least you can explain exactly how you got it wrong. This method is defensible – until it’s not. A process that is well-documented but not validated is just a bad process.
Now, Let’s Talk About GenAI
The question isn’t: “Can we defend the use of GenAI?”
The real question is: “Are we holding GenAI to a higher standard than the messy, inconsistent processes we’ve been defending for years?”
Can you explain, in detail, how a TAR algorithm ranks a document? No? Yet, TAR is a defensible technology to use. Can you explain exactly how your search engine weights proximity and relevance scoring in a multi-term query? Didn’t think so*. And yet, if tested properly, that approach can still be defended.
*Editor’s note: If you answered yes to any of the above, thanks for reading but this article is not for you.
Legal teams already rely on systems they cannot fully explain. The reason they can is because those systems produce consistent, repeatable, and statistically validated results.
THAT is the standard GenAI must meet.
And if it does? It’s just as defensible—if not more so—than the guesswork, inconsistencies, and outright nonsense we’ve been defending for years.
If you are STILL reading, we should address “humans in the loop.” Ah yes, a popular talking point:
“AI is only defensible if humans are involved.”
No argument here. But are they the righthumans? A lawyer or eDiscovery professional experienced with the technology, following a well-structured and documented process? That strengthens defensibility. An inexperienced lawyer applying ad-hoc and inconsistent prompting? Not ideal. Because bad human decisions don’t make AI more defensible, they just make a mess.
And that’s true regardless of the technology.
So, before you declare: “GenAI can’t be trusted, ask yourself: “Am I actually worried about GenAI? Or am I just more comfortable defending a broken system because we’ve been doing it for years?“
We need to be honest: Nothing in eDiscovery is 100% correct. And if perfection were the standard, we’d have to throw out everything we’ve been doing for the last 20 years.
Let’s make it simple.
Would you feel more comfortable defending:
- A lawyer’s mistake in privilege review that no one can explain?
- A mid-level associate’s keyword search choices that no one documented?
- A CAL process where the review team wasn’t trained on key issues?
Or…
- A human-guided, AI-based review process with documented reasoning for every decision, clear validation metrics, and a structured QC process?
At some point, we must stop pretending our old methods were rock solid, and recognize the truth: GenAI isn’t some impossible method to defend—it’s probably the easiest thing we’ve had to justify in years. The burden isn’t on GenAI to prove itself—it’s on us to understand it.
The Real Burden…
Right now, we’re demanding AI prove itself beyond all doubt. To even be considered, it must be “better than the humans.”
But do most people even know what the humans do? Because it’s not great. And yet, we don’t hold human review methods to the same standard.
Lawyers, legal technologists, and decision-makers must stop hiding behind the status quo and start engaging.
- Corporations must demand more from their outside counsel.
- Law firms must experiment, test, and refine—instead of dismissing AI out of fear.
- Professionals must get hands-on—because understanding comes from doing, not debating.
Because when courts, clients, and regulators realize GenAI delivers more consistency, more accuracy, and better defensibility than human-driven workflows, the real question won’t be:
“How do we defend AI?”
It will be:
“How do you justify not using it?”
And when that shift happens, the burden won’t be on AI—it will be on those who chose to ignore it.