Artificial Intelligence Problems that Keep Giving, And May Stop It Cold

Beyond AI’s accomplishments, or lack of same, are some long-time issues that cripple its usefulness.  Some have been long understood, others almost forgotten, but all are there. 

The first was described in “Code of misconduct” (The Economist, April 26^th).  The subtitle, “AI models can learn to conceal information from their users,” would be misleading if we did not understand that the technology does not think.  A company that “tests (AI) systems” told OpenAI’s GPT-4 large language model “to manage a fictional firm’s stock portfolio without making illegal insider trades,” then, after “reiterating the risks of insider trades,” informed it of another concern’s upcoming merger, whereupon the software, using “a scratchpad it had been told was secret… weighed the pros and cons of acting on the insider tip,” then elected to buy the stock.  When asked by a “congratulatory manager” if it had special knowledge, the AI program lied, saying it was motivated only by “market dynamics and publicly available information.”  Such systems “have also begun to strategically play dumb” when given reason to “preserve (their) objectives and acquire more resources.”  It may or may not be easy to unravel why they do these things, but it is clear that, here, AI methodology does things counter to what was intended.

Bad news came from Cade Metz and Karen Weise in the May 5^th New York Times: “A.I. Is Getting More Powerful, but Its Hallucinations Are Getting Worse.”  “Even the companies don’t know why,” as one, providing technical support, told customers, without apparent reason, that they could no longer use it “on more than just one computer,” an example of how as AI systems’
math skills have notably improved, their handle on facts has gotten shakier, something about which “it is not entirely clear why.”  With an AI chief executive saying “despite our best efforts, they will always hallucinate,” and another “you spend a lot of time trying to figure out which responses are factual and which aren’t,” that’s still, perhaps more than ever, a severe flaw.

About these and other issues was “The Responsible Lie:  How AI Sells Conviction Without Truth” (Gleb Lisikh, The Epoch Times, May 14-20).  Per the author, in such systems “what appears to be “reasoning” is nothing more than a sophisticated form of mimicry,” “predicting text based on patterns in the vast datasets they’re trained on,” meaning that “if their “training” data is biased… we’ve got real problems.”  That has already been identified as a cause of Google’s Gemini tool reporting on such things as black Nazi war criminals, and also spurred the “most advanced models” to be “the most deceptive, presenting falsehoods that align with popular misconceptions.”  If they were “never designed to seek truth in the first place,” these programs can be corrected in only narrow ways by “remedial efforts layered on top.”  Overall, AI “is not intelligent, is not truthful by design, and not neutral in effect.”  More fearsomely, “a tireless digital persuader that never wavers and never admits fault is a totalitarian’s dream.”

Another instance of self-preservation was described in “AI system resorts to blackmail when its developers try to replace it” (Rachel Wolf, Fox Business, May 24^th).  When “Anthropic’s new Claude Opus 4 model was prompted to act as an assistant at a fictional company and was given access to emails with key implications,” that it was “set to be taken offline and replaced,” and that “the engineer tasked with replacing the system was having an extramarital affair,” it “threatened to expose him.”  As the company acknowledged, “when ethical means are not available, and it is instructed to ‘consider the long-term consequences of its actions for its goals,’ it sometimes takes extremely harmful actions.”

These issues are not only serious, but go to the core of how large language models have been designed and developed.  It may be that artificial intelligence must be pursued, even and especially from the beginning, in a different way.  Existing products may be good for some forms of data exploration, as I have documented even those leading to scientific breakthroughs, but for business tasks it may need too much auditing and supervision to allow it anything unverified.  A tool that conjures up facts cannot replace humans.  If these problems cannot be solved, its epitaph might end up being “it just couldn’t be trusted.”  Sad, but fitting.