Latest / MBA Training AI / Fair lending and AI credit models: how to defend one

Transcript

Fair lending and AI credit models: how to defend one · MBA Training AI

0:00
Host: Welcome to the Leaders Insights Podcast.
0:07
Host: Today's episode, Building Credit Decisioning Models That Survive Fair Lending Scrutiny.
0:13
Host: A single fair lending violation can wipe out every efficiency gain your fancy AI model
0:18
Host: ever produced.
0:19
Host: So why is every bank racing to automate credit decisions anyway?
0:23
Expert: Because the upside is real.
0:25
Expert: A model can score a loan application in seconds instead of days, and it can approve people
0:30
Expert: a human loan officer would have never looked at twice.
0:33
Expert: The problem is that the same model that expands access can quietly build a bias engine, and
0:39
Expert: regulators have gotten very good at spotting it.
0:41
Host: Let's back up.
0:43
Host: When you say fair lending scrutiny, what specifically are the regulators looking for?
0:47
Expert: Two flavors.
0:49
Expert: First, disparate treatment.
0:51
Expert: That's when you deliberately treat protected groups differently, like coding race into
0:55
Expert: the model.
0:56
Expert: Nobody's dumb enough to do that anymore.
0:58
Expert: The real danger is the second flavor, disparate impact.
1:02
Expert: A neutral-looking rule that still hits a protected group harder.
1:06
Expert: Your model never sees race, but it uses zip code.
1:10
Expert: And zip code in America is basically race with a mailing address.
1:14
Host: So the model doesn't need to know someone's race to discriminate against them.
1:18
Expert: Exactly.
1:20
Expert: That's the trap execs miss.
1:21
Expert: They think, we removed the protected variables.
1:24
Expert: We're clean.
1:25
Expert: You're not.
1:26
Expert: The model finds proxy, Zog.
1:28
Expert: A proxy is a stand-in variable that correlates with the thing you removed.
1:32
Expert: Shopping habits, the name of the college, the phone carrier.
1:36
Expert: Strip out race, and the model reconstructs it from crumbs.
1:39
Host: That sounds unfixable.
1:41
Host: If the model will always find a proxy, how do you build one that survives an audit?
1:45
Expert: You test for it relentlessly, and you document everything.
1:49
Expert: The core technique is something regulators actually endorse.
1:52
Expert: Yee, you run what's called a less discriminatory alternative search.
1:57
Expert: You take your model, and you ask, is there a version that's slightly less accurate,
2:02
Expert: but meaningfully fairer?
2:03
Expert: If one exists and you didn't use it, you're liable.
2:07
Expert: If you searched, found it, and adopted it, you've got a defense.
2:11
Host: Who's actually doing this well?
2:13
Expert: Upstart, the consumer lender, publishes fairness testing on its models and has taken it public
2:18
Expert: with regulators.
2:20
Expert: They report expanded approval rates for Black and Hispanic applicants versus traditional
2:25
Expert: scoring.
2:26
Expert: Google AI has pushed tooling for this too.
2:29
Expert: Their fairness libraries claim they can cut disparate impact substantially.
2:34
Expert: But worth noting, Google sells the software, so cross-check their numbers against independent
2:39
Expert: work.
2:40
Expert: MIT Technology Review ran a good piece showing these tools help but don't eliminate the
2:45
Expert: problem.
2:46
Expert: A bad data pipeline poisons everything downstream.
2:50
Host: Say the CFPB, that's the Consumer Financial Protection Bureau, shows up at your door.
2:55
Host: What actually saves you?
2:57
Expert: Documentation.
2:58
Expert: Boring, unglamorous documentation.
3:01
Expert: You need a model risk file that shows every variable, why it's there, what business reason
3:06
Expert: justifies it, and proof you tested alternatives.
3:10
Expert: The legal standard is business necessity.
3:12
Expert: You have to show the variable predicts default, and there's no less discriminatory way to get
3:17
Expert: that signal.
3:19
Expert: If your answer is the model liked it, you're already writing the settlement check.
3:23
Host: How big are those checks, realistically?
3:25
Expert: Trinity Bank, sorry, Trustmark, settled a redlining case for around $9 million a few
3:31
Expert: years back, and that's on the small end.
3:34
Expert: The DOJ's redlining initiative has pulled over $150 million in settlements.
3:40
Expert: Their efficiency gain from faster underwriting is maybe a rounding error against that.
3:44
Host: Here's the uncomfortable one.
3:46
Host: Isn't the honest move to just keep a human in the loop and skip the AI risk entirely?
3:51
Expert: No.
3:52
Expert: And this is the part people get backwards.
3:55
Expert: Human loan officers are wildly biased.
3:57
Expert: Decades of data show that.
3:59
Expert: A well-tested model is auditable in a way a human's gut never is.
4:04
Expert: You can't subpoena a loan officer's subconscious.
4:07
Expert: The model at least leaves a paper trail you can inspect and correct.
4:10
Host: So the AI is riskier to bill but easier to defend.
4:13
Expert: If you build it right, the negligent version is worse than a human.
4:18
Expert: The disciplined version is better and provable.
4:21
Host: Give me the one thing an exec should do Monday morning.
4:24
Expert: Ask your data science team a single question.
4:27
Expert: Show me the less discriminatory alternatives you tested and rejected.
4:31
Expert: If they go quiet, you don't have a model.
4:34
Expert: You have a lawsuit with a training accuracy score.
4:36
Host: Sources for today's episode, TechCrunch AI, The Decoder, Google AI, Vendor, AI Lab,
4:43
Host: KD Nuggets, MIT Technology Review, ours Technica AI.
4:47
Host: That's it.
4:48
Host: The AI decision tools are live at MBA-training.com.