A Golden Age for Bullsh*t Detection

The Companies That Win with Gen AI Will Be the Ones That Can See Through It

In business as in life, one of the great reframes involves learning to see every challenge as an opportunity. So let me welcome you, belatedly but enthusiastically, to the golden age of bullshit that is also a golden age of bullshit detection as a differentiating strategic capability.

A note on terminology here: I’ve never liked using the word “hallucination” for inaccurate or erroneous output from an LLM, and I like it even less as more companies and products push people to think of their Gen AI tools as interns, co-workers, or partners. If an intern or junior colleague handed you a report with bogus market data and citations to nonexistent sources, would you charitably describe those problematic aspects as hallucinations? I doubt it. More likely: Mistakes, errors, failures. Or maybe just bullshit.

In the discourse around Gen AI, plenty of analysts have opted for the term “bullshit” – specifically in the sense that the philosopher Harry Frankfurt used it to describe statements made with an indifference to their truth or falsehood and according to a logic that prizes things other than veracity. Seems about right, and I’m following that lead here because I have in mind something larger and more insidious than strictly inaccurate outputs returned by ChatGPT & co.: I’m thinking of how we humans represent, utilize, and operationalize AI-generated outputs – accurate or otherwise.

No one needs to be told to remain on guard against unreliable outputs from their LLMs at this point. We’re not batting a thousand on that one as a species, but basic vigilance against nakedly false information should be table stakes. Human in the loop and all that. My concern here is with what happens when the human moves forward with information that isn’t wrong or inaccurate but also isn’t quite what it appears to be – when the information is presented as knowledge but is actually just bullshit content.

In higher education – where the stakes are most obvious, this has quickly become a nearly existential challenge. Universities have scrambled to define systems that recognize the distinction between knowledge and BS content, reward the former, and incentivize actual learning – rather than the production of content that may address a prompt or question correctly but doesn’t ultimately signify anything of value. Now, sure enough: Amid this crisis, an opportunity has indeed emerged – and not just the opportunity to think deeply about what higher education should look like in a changing world and why. It’s also a boom time for… blue books and other solutions that can recognize and reward work that’s actually aligned with the purpose of the university, an institution which exists – at least in part – to confer a meaningful credential that signifies some level of knowledge and achievement to the job market, etc.

If the university can’t reliably detect and guard against bullshit (especially the factually correct bullshit), then the credential is devalued, and the whole enterprise finds itself on very shaky ground. Companies face different versions of the same problem, but fundamentally, their success and survival in the gen AI-enabled future will depend on the ability to reliably detect bullshit (especially, the factually correct bullshit).

Unlike the output of work performed by university students, which is part of the learning and credentialing process and serves largely to establish and indicate proficiency, much of the output generated by knowledge workers in a company at some point needs to be operationalized. The business plan has to be executed. The mockups have to go to production. The code has to be shipped and needs to run at scale.

This is where any gap between “looks good, sounds good” content and actionable knowledge backed by real capability and experience is going to come to light. The vibe tax comes due. The promising project hits a development bottleneck. Superficial capability is exposed. The deadline and budget are blown. Credibility and trust are compromised. And not because the vibed-up, AI-enabled MVP didn’t work, but because it did – well enough to convince the insufficiently critical to buy a little too heavily into their own (just true/functional enough) bullshit.

The companies that hone a capacity for bullshit-detection – not cynicism, but clear-eyed evaluation grounded in epistemic humility, an understanding of models’ limitations, and a healthy respect for what it takes to move from compelling POC/prototype to reliable/scalable product – will be the ones that turn gen AI powers into leverage rather than risk. They’ll be able to budget for the vibe tax and plan for the hard work that can’t be handwaved while keeping the momentum and capturing the value of accelerated exploration and discovery.

The companies that lack – or lose – that strategic capability (and discipline) of discernment may find their most exciting ideas stumbling down the stretch, beset by foreseeable difficulties, just beyond the ability to deliver, and generally mired in, well, bullshit.

@Jeffrey