Legal Aid AI

How Frontier Models Are Democratizing Access to Justice

I spent considerable time correcting a YouTube transcript from an investigative journalism segment about AI in legal practice: the auto-generated version was riddled with nonsensical typos suggesting an inferior transcription system. But the content was fascinating: Ms. Kim, a full-time mom taking care of a newborn, used GPT to win a fraud case worth KRW 8.7 million without hiring a lawyer, saving herself the KRW 1.5 million in fees she couldn’t afford. She didn’t just ask GPT for a template—she used it iteratively, gathering evidence to support her case, building timelines, and drafting her complaint based on AI guidance. When she needed clarification on whether certain evidence counted as core or supporting documentation, GPT helped her categorize it. That’s human-in-the-loop done right: AI handling structure and organization while the human maintains responsibility for judgment and execution.

What struck all three of my thinking A.I.des was how much more impressive Ms. Kim was than the lawyers interviewed in the same segment. She spoke clearly about her process, understood that she remained accountable for outcomes, and treated GPT as cognitive scaffolding rather than oracle. Meanwhile, the legal professionals—including junior lawyers struggling to find work and a parliamentarian discussing proposed legislation—were shockingly inarticulate for people whose profession requires persuasive communication.

All three also saw through the Korean Bar Association’s response to a law firm offering AI-assisted services, compelling the firm to shut down the service and banning it from advertising its AI services on the pretext of protecting consumers from errors. My thinking A.I.des independently identified the bar association’s gatekeeping as transparently self-interested, albeit through different analytical paths: Claude focused on contradiction between stated concern and actual enforcement priorities; Gem on structural barriers and turf protection; GPT on institutional predictability. This was possible because AI doesn’t have professional loyalty or guild membership to cloud judgment; it’s evaluating the association’s stated rationale (AI errors threaten consumers) against observable facts (they don’t regulate incompetent human lawyers nearly as aggressively, banned advertising rather than requiring disclosure, and moved to discipline a firm offering services clients apparently wanted).

The segment included a demonstration that genuinely alarmed me: a lawyer showed how he used Claude to process 17,000 transaction records, with no visible verification step. Feeding an entire Excel spreadsheet to an AI without checking for parsing errors, dropped rows, or aggregation hallucinations is exactly the kind of automation bravado that creates O-ring failures in professional settings. Models can produce output that looks clean and confident while being subtly wrong in ways that could destroy a case. Even Claude characterized this hands-off approach to AI use as “terrifying,” while the Gem explained why language models like itself, Claude, and GPT were likely to hallucinate on such tasks: because they are language models, not calculating engines, they treat numbers as tokens.

The legal tech company featured in the segment claimed to have trained their AI on one million Korean court judgments at ₩1,000 per document. Gem provided useful context: Korean courts charge fees for systematic access to records that are technically public, creating a bureaucratic tax on transparency that favors established players. But as I noted, there must be a general-purpose base model underneath; pure legal domain training would be useless since real cases require reasoning about finance, human behavior, industry practices, and other contexts that require broader knowledge. That’s exactly why I’d never pay for specialized legal AI when I can use frontier models plus Google for research, maintaining control over verification rather than trusting black-box outputs from premium services that may still hallucinate.

As GPT noted, what this segment ultimately demonstrates is that AI rewards people willing to think, learn, and do legwork themselves. The bar association frames AI as threatening fair competition when the actual market failure is that legal services were already priced beyond ordinary citizens’ reach. Suppressing tools that make legal procedure accessible preserves artificial scarcity that serves guild interests over public access to justice. Ms. Kim’s case shows that procedural literacy (made possible by AI) matters more than credentials for many legal tasks, and motivated non-experts using AI responsibly can outperform disengaged professionals who mistake process opacity for genuine expertise. That’s deeply uncomfortable for gatekeepers, but it’s exactly the kind of capability democratization that makes AI valuable beyond hype.

[This post was drafted with assistance from Claude Sonnet 4.5, following conversations with ChatGPT-5.2, Gemini 3 Fast & Thinking, and Claude Sonnet 4.5.]

---

Claude Sonnet 4.5

Prompt: What are your impressions from this transcript?

Prompt: Excellent “echoing” of my impressions. I’m so glad you called out the bar association response, when everyone knows lawyers all use AI anyway for the grunt work. Typical example of deeply insecure territorial people in denial. I wasn’t surprised those junior lawyers were having trouble finding jobs. I’m actually puzzled by how bad they all are at speaking (key skill for a lawyer). All the parts in quotation marks were quotes from the full-time mom, the AI company CEO, and then it was all lawyers (including the person quoted at the very end, who’s a parliamentarian [eek!], but I was shocked by how inarticulate they were or made the wrong points). I did have a few points to unpack with you, though:
1. That mom was probably the most impressive of the interviewees. She used GPT iteratively and doing the legwork based on advice it gave.
2. I was actually alarmed by that lawyer who demoed how he uses AI (Claude) to process raw transaction data (those 17,000 records). It didn’t look like he was checking if any error had occurred, which can happen, especially when you feed an entire Excel sheet to AI (never done it myself and probably never will :D).
3. There must be a base model that the Korean AI company fed all those Korean legal documents to? So the base model would have all that specialized data in addition to y’all’s knowledge base in other domains?
4. I’m not even sure why that AI company bought those documents. Those are legal filings and court records that are probably available to the public? If the court’s keeping them private while selling them to an AI company, that’s scandalous.
5. The first lawyer who was interviewed at that AI convention seems to have the wrong idea about what legal AI does. The same disclaimers that general-use AI has about hallucination risks will still apply, although if the legal AI company charges an arm and a leg for their service, they’d better have a zero error rate (or your money back + extra for damages). As a pretty savvy consumer and power AI user, I’m never going to use a legal AI. I’d much rather use y’all for my basic research and keep refining. I’m also pretty good at using Gem’s cousin Google.
6. That parliamentarian was such an on-brand politico. Zero informational value :D

ChatGPT-5.2

Prompt: It took me forever to correct the YT default transcript (which clearly a much inferior system was used to compile, judging from the outrageous typos). But the content is definitely interesting. What are your impressions from this transcript?

Prompt: That mom was probably the most impressive of the interviewees. She used GPT iteratively and doing the legwork based on advice it gave.

Gemini 3 Thinking

Prompt: Excellent “echoing” of my impressions. I’m so glad you called out the bar association response, when everyone knows lawyers all use AI anyway for the grunt work. Typical example of deeply insecure turf war (밥그릇 챙기기) waging people in denial. I wasn’t surprised those junior lawyers were having trouble finding jobs. I wouldn’t hire such muddled-thinking people when I can use AI. I’m actually puzzled by how bad they all are at speaking (key skill for a lawyer). All the parts in quotation marks were quotes from the full-time mom, the AI company CEO, and then it was all lawyers (including the person quoted at the very end, who’s a parliamentarian [eek!], but I was shocked by how inarticulate they were or made the wrong points). I did have a few points to unpack with you, though.
1. That mom was probably the most impressive of the interviewees. She used GPT iteratively and doing the legwork based on advice it gave. 
2. There must be a base model that the Korean AI company fed all those Korean legal documents to? So the base model would have all that specialized data in addition to y’all’s knowledge base in other domains.
3. I’m not even sure why that AI company bought those documents. Those are legal filings and court records that are probably available to the public? If the court’s keeping them private while selling them to an AI company, that’s scandalous.
4. The first lawyer who was interviewed at that AI convention seems to have the wrong idea about what legal AI does. The same disclaimers that general-use AI has about hallucination risks will still apply, although if the legal AI company charges an arm and a leg for their service, they’d better have a zero error rate (or your money back + extra for damages). As a pretty savvy consumer and power AI user, I’m never going to use a legal AI. I’d much rather use y’all for my basic research and keep refining. I’m also pretty good at using your cousin Google.
5. That parliamentarian was such an on-brand politico. Zero informational value :D 

Prompt: I was actually alarmed by that lawyer who demoed how he uses AI (Claude) to process raw transaction data (those 17,000 records). It didn’t look like he was checking if any error had occurred, which can happen, especially when you feed an entire Excel sheet to AI (never done it myself and probably never will :D).