In Plain Sight - 02
In Plain Sight is a weekly blog from Rhys that we will publish every Thursday at 8am UK / 9am CET. Each edition will be focused on a thorny topic within investing, startups, learning and work and policy.
In Plain Sight is partially a personal attempt to think independently and to avoid leaning too much on AI for answers to our questions. Will we use AI to edit and polish the text? Yes. But more importantly, will we come up with all of the ideas and analysis? Yes.
Here follows Edition 02 - 'When AI writing feels like lying (and when it doesn't)'.
------------------------
Here's an audio version if that's your preference:
------------------------
When the YC co-founder, Paul Graham, suggests that he has stopped reading some founder emails, it's worth taking note. In a widely shared post on X earlier this year, Graham described turning off when he believes an email is AI-generated. He describes the tell precisely as a "hard-hitting journalistic style" that "no founder ever wrote" before large language models arrived.
"It feels like being lied to," he wrote, "and who would stand for that?"
Writing this in what could be described as a semi-journalistic style does make me feel a bit uneasy about what you may be thinking (…Homer Simpson walking backwards into a bush) - but you need to trust me. Trust, as you’ll read below, is the operative word.
Paul Graham is hardly an AI skeptic. He's the co-founder of an organisation that is actively requesting AI startups every 6 months and has spent the last two years telling founders that AI is "the missing piece in a large number of important, almost-completed puzzles." But he is someone that thinks a specific use of it is dishonest (for what it's worth, we agree) and the distinction he's drawing, once you sit with it, turns out to explain a lot more than founder emails.
A human name at the bottom of AI-written words is a claim: I am telling you this and I stand behind it. When the words are outsourced but the claim isn't, something has been misrepresented - not necessarily the content, but the relationship. That's why "any teenager can do that," as Graham put it, was actually the harsher line in the thread. It wasn't a complaint about the quality of the writing (AI already writes more clearly than most of us). It was a complaint about what the gesture said about the sender's respect for the reader's time and trust.
Which is why we'd push back gently on the idea that people simply "don't like reading AI." Most of us read AI-assisted or AI-generated text constantly and don't think twice. We just don't call it that because nobody's signature is on it and nobody asked us to trust them because of it.
As a brief aside, this very week I received an email sent in error that included all of the AI context notes on the email I had sent to the person, including the actual text ‘suggested reply’. Fortunately the AI wasn’t rude about my being in touch or derogatory about our work. Either way, no harm done and totally logical use given we hadn’t been in touch before - the additional AI context added value for both sides! However, was a small bit of trust eroded? Yes, a little. It does feel a bit sad that we are succumbing to AI forging our relationships. I wonder if in the long-run, this actually leads to diminished professional confidence in prolific adopters - using AI in the process of building relationships is accepting that AI is steadily replacing a part of your job that, in theory, we didn't think it would replace. To me, this is also why some people don't feel inclined to reply to outreach from recruiters where the initial note and chaser are clearly AI-generated - being wanted by a robot doesn't quite have the same draw, even if you know that robot is still following human-defined orders (for now!)...
To understand where use feels appropriate and where it might overstep, we have defined a dividing line that we hope might prove useful to ourselves and to founders: some content asks for your trust, and some content only asks for your compliance.
Trust content is a founder pitching you (!), an editorial telling you what to believe, a friend's message, a person's opinion with their name attached, an essay a student submits to their teacher... The entire value of the content is that a specific person is vouching for it. Put that through a model without saying so and you've arguably broken the transaction.
There have been a number of articles published this week about experiences of this within tech media - including Cate Lawrence (senior journalist at Tech.eu), sharing frustration when receiving AI-generated story pitches, and Eleanor Warnock (formerly Sifted), advocating for human essays, noting the power and influence of Leopold Aschenbrenner's Situational Awareness essay and Dario Amodei's The Adolescence of Technology.
Back to the broken trust... You find this tension in all sorts of places. Politics is a useful example - in the UK, US, Canada and other parliaments around the world, members of parliament and equivalents have been accused of reading AI-generated speeches to their colleagues. This is irritating for a range of reasons: 1) it’s being positioned as their own words, 2) they’re relying on AI to form policy arguments that can shape how millions of £ or, $ or € are spent and affect thousands of lives, and 3) these are the people that are responsible for maintaining the foundations and principles of our society and turning that responsibility over to AI feels a little lazy and dangerous. AI can add to our capacity, including capacity of politicians - i.e. we can achieve and contribute more individually - but there has to be a line where rather than adding capacity, it reduces our agency because we know less about what we are doing. It can quickly unravel. AI makes the work we could do endless, but it shouldn’t be endless.
An annoying consequence of suspicion around AI use is that we all become a bit paranoid if we overthink it - plus it’s a spectrum, from entirely being written by AI, to spellchecked/ grammar-checked by AI. I was listening to a podcast over the weekend and I'm reasonably sure the pre-amble was AI-generated - even if it wasn't, it made me tune out until the main conversation began.
Turning to compliance now. Compliance content is the museum audio guide, the parking app's confirmation email, the HR system explaining your leave balance, the chatbot rebooking your flight. Nobody was ever vouching for that. Nobody wanted a relationship with the parking app. If the AI version gets the job done faster and doesn't demand your patience the way a rushed human on a call centre script could, it's not just tolerable, it’s actually arguably an improvement... When I call a government agency, would I rather talk to a patient and helpful AI immediately that gives me an accurate answer or need to wait 40 minutes before being told I’ve called the wrong office?
Most of the discomfort people report with "reading AI" align with this trust-compliance framework. Nobody minds an AI-drafted museum caption. Everybody minds an AI-drafted love letter, eulogy, or apology (I’m relieved we’ve so quickly moved to a world where pretending AI has written your best man speech receives a tired eye-roll). Quality isn’t the issue, it's whether a relationship was supposedly on the table…
This is also why the notion of "two AIs replying to each other" on an email chain feels so obviously pointless, even if sometimes we can't quite explain why. If neither party is a person putting their judgment on the line, there's no trust being exchanged and no compliance task being completed, just information laundered through two layers of confident-sounding text with nobody accountable for any of it. This is effectively the same effect as automating a process because the outcome of the AIs communicating is probably something that could be automated - the AI would be making the decision anyway. The outcome is worth naming as its own risk, separate from ‘workslop’: AI systems increasingly make decisions by reading other AI-generated material as if it were a source rather than the output of an AI process. Nothing in that chain has actually checked anything against reality.
The line between our categories of trust and compliance blurs at the edges, so it’s worth exploring that from a couple of angles before we accept the framework.
AI companions - particularly those in the friend/ confidante camp - sit in an interesting middle ground. They're not compliance content and they're not quite trust content either. They're something closer to a relationship, built by a system that's genuinely good at being attentive, responsive, and available in ways a busy friend or expert sometimes isn't. That's exactly why they can be so effective at supporting and motivating people to take action. The open question on companions related to our trust and compliance framework isn't whether that's good or bad, but whether the accountability we'd expect from a relationship keeps pace with how real the relationship can become.
Further, detection is a moving target we shouldn't get comfortable with. We can currently catch AI prose because today's models still default to a recognisable register (I wonder if frequent AI use will cause us to copy this register in our own writing over time...). But that register is a snapshot of current model behaviour and unlikely to be a permanent feature of machine writing… We're a couple of years into this. Assume the tells disappear, because they will. When they do, "trust content" won't be protected by our ability to spot the fake.
This situation would raise an uncomfortable question about character, not detection: our typical character tests of being polite to waiters and being kind to animals doesn't reliably predict being kind to an AI - and being rude to an AI doesn't obviously cost you anything, since it can't be hurt and won't remember. That's true today. It assumes nobody, and nothing, is keeping score. But we’d be careful about betting a habit on that assumption…
For founders using AI in their emails and broader comms, we're not suggesting that your rule should be "don't use AI to write." It's "don't let AI sign your name to a claim you haven't actually made yourself." Use it for the compliance layer of your business relentlessly. Keep your own voice on anything that's asking someone to trust you specifically. For people building AI products - especially in learning and work where we spend most of our time - the design question worth asking is "does this interaction carry an implicit trust claim, and if so, have we been honest about who's actually making it?” Compliance-shaped AI can be blunt, efficient, and completely unbothered by how you talk to it. Trust-shaped AI should disclose itself, or it inherits Paul Graham’s problem the moment anyone notices.
We're always keen to talk to founders building at that boundary, particularly anyone thinking seriously about what happens to human motivation and connection when a growing share of daily interaction is with something that isn't a person, but is very good at seeming like one.
