Item — HN Reader

The problem we're seeing across many professions is AI output is not getting vetted by knowledgeable people, whether it's an experienced analyst, senior engineer, expert attorney, or the resident physician. At best they skim, at worst they don't even see it at all before it's published, pushed to production, distributed to clients, or submitted to the court.

In many cases the skills are available in house to do the necessary vetting, but these people are already overwhelmed with their existing day to day.

Anyone remember that item a few months back about Amazon now having senior engineers vet generative AI output (https://news.ycombinator.com/item?id=47323017)? I had to LOL when I read that. These folks are already slammed. And the idea that Amazon would allow human bottlenecks to multiply across projects and underlying infrastructure development is ridiculous.

18 comments

Part of the problem: you get given a complete document to review after it's been fully baked.

I'm pushing the need for basic engineering principles across whole organisations.

You wouldn't give an engineer 1000 lines of code to review without the original spec of what you're trying to achieve for context (at a minimum, ideally the reviewer was in the room when the work was introduced, and has full context).

So, these docs, they're given as an all or nothing.

Do you push back on the 39th metric that is defined to the utmost detail? Or just resign yourself to the fact that it is what it is?

A one (6 is the goto if we're talking Amazon?!) pager.. "this is what I am proposing" at least gives the skeleton of the idea to push back at the general shape of the idea, refine it, before all the emotional investment of your precious report being complete.

Y'know.. the traditional product running through the spec in a SCRUM* environment.. the engineers doing proper code reviews..

* Yes SCRUM is dead, but that's another thing.

> Part of the problem: you get given a complete document to review after it's been fully baked.

Not fully baked, worse: made to sound confidently correct, orthogonal to its actual correctness.

Like the fake food they make for commercials. Looks great on TV.

So many people get buried in tech debt due to constantly choosing "let's be productive now and iterate later". Then later never comes.

This was already proof that, when a company is given the chance of getting a reward now and putting in the effort later, they will take the reward and postpone the effort indefinitely.

AI now offers "generate now, review later". Fill in the pattern.

We're cooked and it's not due to AI, it's due to the fundamentals of an economy and society that only sees the short term.

Is VibeDebt a term yet?

I've had this situation and basically just had to throw out stuff that was written because its completely terrible/wrong. Either start again or just give up.

As an attorney, I feel like vetting AI output takes longer than just doing it from scratch, let alone versus just using a traditional form.

With AI, I have to read through everything, often explain why it's wrong, and then rewrite everything anyways. I mean, I get way more billables, but I think it's symptomatic of how AI loses its advantage of being quick and accessible to those who don't understand the subject matter.

>As an attorney, I feel like vetting AI output takes longer than just doing it from scratch, let alone versus just using a traditional form.

This is my issue with AI.

In the type of work i do the work needs to be precise down to the context of how individual words are used.

Having AI pump out 20 pages of content but then me having to go through the 20 pages word by word, cross checking references and prior statements is going to take a long time.

Not to mention I didn’t write it, so my brain doesn’t already know what’s been written so it takes several passes to confirm its complete and that it all fits together.

I find it easy to just write it myself use AI for the more menial tasks like logic check, completeness checks, etc.

One task AI was very useful in was where we wanted to understand the gaps in a submission relative to a process requirement document. We didn’t care if the output was 100% complete or perfect, we just wanted a few examples.

Being able to input a couple 300 pages documents and have AI spit out a dozen examples in 30 seconds was a huge time saver.

But that was a non-critical task.

Fact-checking and editing a mediocre piece of writing be way harder than writing from scratch. Proving that something isn’t true or can’t be substantiated is hard work, and so is arguing that a word choice is subtly inappropriate.

And making a ton of corrections to a document everyone was hoping was ready to go is never fun politically.

Another attorney here. I understand your plight. But I can't believe law firms are sending out briefs and opinions without carefully checking all of the citations. I mean, even when Lexis or Westlaw identifies an (actual) case on point, you still have to check if the case has been overturned, whether it is truly on point, or if it can be distinuished from your case. So even if the cited case is not a halucination, someone would still have to read and analyze the cited case in the context of the present case.

>> But I can't believe law firms are sending out briefs and opinions without carefully checking all of the citations.

Update your priors: https://www.damiencharlotin.com/hallucinations/

It's not really any different in programming. Like if you have a well structured code and want to do a clear refactoring across it and you know what to expect, it can speed things up. But if it's generating any significant (and relatively complex) new code, you have to go through the whole thing manually again and then you find out you have to fix way to many things and get bogged down in different paths the AI didn't do correctly.

Of course, it's pretty much impossible to hear a dissenting point of view today and everyone is going crazy on these drugs. I might be hilariously wrong but I think this is the best time to start a software company.

Youre not wrong I believe.

I think its the perfect time to be contrarian - think about it. If youre wrong - So what? The world will have changed for everyone in the field. If you are right? You stand to be positioned to win big financially whilst everyone elses brain is rotting away.

Be afraid, be very afraid:

"AI Hallucination Cases" - https://www.damiencharlotin.com/hallucinations/

You can also feed the document or source file to another frontier-level model, ideally two others, and tell it to vet it aggressively. The goal is to goad the models into erring on the side of false positive findings rather than potentially missing true positives.

I find that if Gemini Pro agrees with Claude Opus 4.8 and GPT 5.5 on something, it's almost certainly correct at a level where I wouldn't be likely to catch any errors myself.

How do you use it, as in, hey, write a doc about this, or do you iterate more like a conversation?

I do the second approach for coding with smallish steps and the output is fine

To be honest, I don't use it at all professionally except I guess transcription or search engine summaries to point me to actual documents. I mostly come across it when a client makes a contract or legal memo on his own and sends it to me to review. If I'm drafting a lease or a trust or whatever, I already have forms that cover everything they need, have been used thousands of times, and can be easily tweaked as needed cuz I half have them memorized at this point.

I can't imagine using them for a court filing. Those are either so short that they're trivial or they require painstaking research, precision, and often citations for every single sentence.

This is the realization I had too. We had a manager update a policy at our org. He just shit it out through AI. It had tons of mistakes, people who read it had questions. Not only did it have mistakes it was causing people to do things in a way that added a manual step when an automatic process existed. Then the engineer VP commented on it asking the original author what its about who then had to bring it back up to the attention of the manager who made the first change.

It wasted many people's time, probably an order of magnitude of time wasted (and money) than if the initial person put a modicum of effort into making it right in the first place. Instead they hand it off to their life partner claude and just assume its good enough.

It's to the point where I am feeling insulted when I get ai slop like this from people. If I am expected to perform at a high level then I expect that at the very minimum the slop throwers will proof read their slop.

Ugh same. I take pride in delivering software that is bug free, performant, and to spec.

When our product managers just send us AI generated JIRA tickets that are extremely long but contain no actual details and tons of irrelevant or wrong content I get extremely frustrated and are seen as not a team player. At that point I’d rather not have the product managers present in the process.

I have experienced this several times lately when writing software with claude/codex. Sometimes vetting and steering the agent takes longer than it would have taken me if done manually. Sure you can just decide not to vet the output and go into full vibecode, but agents tend to do a lot of dumb things (such as not deleting unused private methods or having temporary variables that are not needed).

In my experience the most effective work pattern for me is using agents to perform research and feedback on high level design, then I write the code manually, then I ask the agent to review the code for potential bugs/issues and fix those. The agents have a much easier time making small changes once the design is 90% there without going fully off the rails and generating slop.

I am working on writing skills to make the agent better but it is a bit painstaking. For example I had to write this inside of a skill because sometimes the agent would just stub out methods and leave TODOs: “always fully complete the requested task before finishing edits unless input is needed”.

I’m against “vibe” anything important, but the fundamental flaw with this reasoning is that unknown unknowns exist.

I can’t cite “from scratch” for something outside of my knowledge but I side LLM training or assisted search.

> AI output is not getting vetted by knowledgeable people

You mean the people they fired and demoralized?

One of the things that "great [wo]men" like about "vibe-coding" (and that includes blindly producing non-code product), is that they, and they alone can now do what used to require the painful process of "passing it to context experts."

Now, the LLM is a "built-in context expert," and they don't need to vet the output anymore.

> Now, the LLM is a "built-in context expert," and they don't need to vet the output anymore.

Serious orgs are going to have to figure out the human layer. It will be needed, no matter how 'hallucination-free' the AI tooling gets. AI will still have some spectacularly bad fuck ups or even worse time bombs that get embedded in a system and don't become apparent until months or years later.

A lot of this will be dumped on existing staff with predictable results as they don't have the bandwidth to do it right. I can envision "output compliance" or "AI QA" becoming dedicated positions at many orgs. It's clearly needed.

> It's clearly needed.

Once the hallucination rate drops below error rate of human workers, it won't be needed anymore.

Let's be honest, how many orgs are really serious? Playing the game of the day for shareholder appeasement is taken far more seriously than whatever the domain experts might think.

> The problem we're seeing across many professions is AI output is not getting vetted by knowledgeable people

The problem is that output sometimes take longer to verify than to create in the first place.

That turns AI into a deeply negative ROI system for many applications.

> In many cases the skills are available in house to do the necessary vetting, but these people are already overwhelmed with their existing day to day.

This is an interesting topic. We treat vetting output the same as doing the work ourselves, but that is not the case.

Doing the work is not the same as reviewing work done by others.

I have heard reports of software engineering companies that have gone full agentic. Their seniors only review stuff written by LLMs and it burns them out, because they have to switch context constantly.

I find this interesting because part of being a senior developer is that you are experienced enough that you won‘t make grave mistakes anymore. This is the case in many professions: you are relied upon to not make grave mistakes.

But those same people are now swamped with stuff that they are not able to review, so they will let a grave mistake slip through at some point.

So they really can‘t trust themselves anymore?

>>The problem we're seeing across many professions is AI output is not getting vetted by knowledgeable people

I am particularly interested in Education and Human Knowledge Management. I have seen the rate of IT training going to zero. Think about specialized training, where if you make a mistake, the consequence of your errors, are talked about on the tv news of the evening.

The whole idea everybody is just planning to save their butt, using these strings coming out of these numeric matrices, while suspending judgement, just shudders me in horror. A bit like those South Asia Airline companies, that were forbidding their pilots from landing airplanes with manual piloting, leading to an increase loss of skills causing some well known disasters...

If well paid consultants cant even bother to check their links...

Also wondering on this whole review process with someone who wrote it with AI. Even if you comment and noted all issues. Do they have skills or willingness to correctly correct it all? And how many times would you need to keep the loop going for error free outcome? Is there even enough calendar time for that?

> In many cases the skills are available in house to do the necessary vetting, but these people are already overwhelmed with their existing day to day.

I think a lot of the time it's just pure laziness. AI gives people a magical "do all the work for me" button and it can bring out the worst in them.

I constantly battle this dichotomy where I care about the work I do but I also cannot possibly care about the corporate model, given 0 ownership of flawed processes across the org and the looming layoff that'll happen any day now.

Some people are given the button and really do not care.

It looks like when you sum up: the cost to generate information using an LLM + the cost to actually verify the information, the result on average is the same as not using the LLM. That does not mean some times it isn’t faster and cheaper. It just means other times it’s slower and more costly. This along with different people’s tolerance for accuracy explains why we see such diverging experiences with it.

So in order to make it pan out the forces at play are trying to make everyone believe we now have to accept wrong, even dangerous results.

But wait, if knowledgeable people have to vet the output, the process will not be 10X faster and you will not be able to fire the knowledgeable people. Therefore, your objection makes no sense. QED.

>> Amazon now having senior engineers vet generative AI output

Software and news are apples and oranges. The software engineer looks at AI code to spots errors and make/suggest corrections. But the software still exists. A doctor vetting a news article ("Owning a Boat Causes Cancer") isn't there to tweak language. The doctor's job is to stop the entire article. It would be like the amazon engineer deleting all the source code and telling the dev team to abandon the project. That will never be a popular task.

It's similar to tokenmaxxing in various companies. Who cares about reviewing the design and the code, just code generation all the way. If it runs it's good

The problem is that a lot of employees across the corporate world are engaged in malicious compliance with regard to AI usage.

If they were going to have employees do the in house vetting, they could just have them write the report.

> the idea that Amazon would allow human bottlenecks to appear across projects and underlying infrastructure is ridiculous.

Why?

Amazon is fairly well known to ruthlessly optimize every process.

So if they're having humans proofread what the AI produces, they must have found that to be necessary.

Yeah, that's what I think too. They aren't going to care about optimizing a process that leads to poor results.

Amazon is not immune to making mistakes

It is harder to check everything then to create a thing without lying in the first place.

If the main job is putting out a report, starting with AI is wrong in any case. What's the value of an AI-generated report, even if experts fix the biggest issues with it? Maybe this kind of report didn't have all that much value before, I don't know. But starting with AI just makes sure it's generic drivel.

> The problem we're seeing across many professions is AI output is not getting vetted by knowledgeable people, whether it's an experienced analyst, senior engineer, expert attorney, or the resident physician.

Yeah probably not for the same reason I left VFX rather than have a lifetime of completely disregarding my own generative creativity and cleaning up LLM-generated bullshit. Fuck that. Double-fuck creating ‘content’ to train the models.

In code, LLMs automate away a lot of the drudgery. I wasn’t sad to avoid spending a couple hours looking up the usage patterns and idioms for some ported library, or do some rote task that didn’t make the project significantly better. In most other jobs, they automate away the only fun part and leave humans with all of the drudgery.

The tech industry has always been arrogant to some extent, but assuming the world of talented professional knowledge workers and creatives would be content to professionally proofread, apply lipstick to pigs, and polish turds is a whole new level of out-of-touch. I’d rather live out of my car and dig through the garbage for bottles with deposits.

> In most other jobs, they automate away the only fun part and leave humans with all of the drudgery.

I don't know about "most other jobs", I feel like this applies strongly to software too. AI is being used to automate the fun stuff

You may view writing code as drudgery, I like doing it

Some people enjoy bookkeeping, too.

I loved solving problems with code but I don’t think it’s controversial to say most people enjoy the novel problems much much more than pulling apart an obtusely designed, non-idiomatic API. The AI is better at the details than the novel problem, so whoever is driving still needs to sort that out. In art, writing, and so-on, it only does the interesting part, but it does it poorly. So for professional use cases, it involves a lot of tedious cleanup of other people’s art that a machine (very economically) amalgamated into a pile of shit, and you’ve no opportunity to do anything interesting. At all.