[PREV - DOCTOROWS_ACCOUNTS]    [TOP]

AI_SNAKE_OIL

                                             March 10, 2026

  "AI Snake Oil" (2024) by
   Arvind Narayanan & Sayash Kapoor             FIRST13
   First paperback edition: 2025

      Reading the preface to the paperback
      edition I note a few red flags--
      the authors are posing as Reasonable and
      Open-Minded.  This could be genuine,
      but in recent years we've seen a lot
      jamming by people coping that pose           But maybe the Moderate
      while looking a lot like stealth             Extremists deserve to be
      advocates for the right.                     thought of as a faction
                                                   of their own.
      I see they identify as programmers
      and I wonder what kind of programmer
      they are-- and how old they are, which
      actually does matter: there have been
      generational shifts among programmer's
      and one suspects they're definitely
      post-Brogrammer era.

From the preface:

  "On the other hand, we are cautiously
  optimistic about generative AI
  applications such as chatbots. ... "

                                         But what follows isn't "cautiously
                                         optimistic", it's a glowing endorsement.

  "Turning to actual generative AI, we
  say in the book that it will be
  useful in some fashion to most
  people whose work is cognitive in
  nature. We are heavy users of
  generative AI ourselves, especially
  for computer programming, an area
  where capabilities are rapidly
  advancing.  AI can generate snippets
  of code, and even entire
  applications, albeit simple ones,
  based on a text description of what       This endorsement gets less impressive
  the code should do. ... For many          when you're aware of how programmers
  programmers, including us, it's hard      can fool themselves about the
  to imagine going back to an era           effectiveness of the latest fads.
  before the availability of AI
  assistance for writing code."             Consider this result from
                                            Joel Becker et al:

                                            "Before starting tasks, developers
                                            forecast that allowing AI will
                                            reduce completion time by
                                            24%. After completing the study,
                                            developers estimate that allowing
                                            AI reduced completion time by
                                            20%. Surprisingly, we find that
                                            allowing AI actually increases
                                            completion time by 19%--AI tooling
                                            slowed developers down."

                                            https://arxiv.org/abs/2507.09089

  "On the other hand the difficulty of
  avoiding pitfalls has also been
  increasing over time as generative AI
  has become much more capable and          So it's become "more capable"
  products have proliferated. For           and yet the difficulties have
  example, there can be subtle bugs in      increased.
  AI-generated code. For now, it is
  *caveat emptor*."                         I don't doubt that there
                                            can be "subtle bugs".
                                            It wouldn't surprise the
                                            hell out of me if they're
                                            can be major bugs.
                                            This can also happen with
                                            human developers, of course
                                            so the queston is frequency
                                            and severity, and whether
                                            you have people on staff who
                                            are competent enough to fix
                                            them, and whether they'll
                                            still be able to after years
                                            of "AI" addiction.

                                                (Though really, that's just
                                                one set of questions.  There
                                                are many others.)

  "... it takes time and experimentation
  to develop a working understanding
  that is tailored to the specific ways     The subtitle for this book claims
  you might want to use AI in your          to be able to *answer* that
  workflows. Given that there can be a      question-- they're supposed to
  significant leanring curve, is            be able to tell you "how to
  generative AI still worth it from a       tell the difference" between
  productivity perspective?"                what AI can do and what it can't.

                                            In this preface, they're backing
                                            off and managing expectations.


 Other issues that might be raised-- (they
 aren't touched on in this preface, at least):

     Will widespread adoption of "generative AI"
     pollute the source data with "AI slop" and
     reduce the quality of output?

     If the massive data centers and their                   (~p. 273)
     current tremendous energy usage are           They claim new regs
     actually necessary, should they be            aren't needed, just
     allowed to exist?                             enforcement of old.
                                                   Okay.
     The big players are still running at a
     big loss, giving away services for
     free in hopes of attracting more              In other words, we're
     interest.  What happens when they             in stage one of the
     switch modes and play for money-- will        Enshittification process.
     Narayanan and Kapoor suddenly be able
     to imagine "going back"?                            ENSHITTIFICATION


--------
[NEXT - FAILED_FUTURES]