Recent, widely discussed research study by Wiles and colleagues of Boston University, "Why you shouldn't treat AI agents like employees," is at least revealing, maybe even shocking to some. In essence, this is what the study involved, summarized (pretty accurately, actually) by Perplexity)AI:
"Wiles and colleagues surveyed and experimented on more than 1,200 managers and executives, mainly in HR and finance roles. In the core experiment, everyone reviewed the same kinds of documents (for example, drafts of job ads or financial memos) with planted errors, but the label on the work varied: some were told the draft came from an “AI tool,” others from a human employee, and others from “ALEX‑3, an AI employee/coworker.” The underlying model texts were identical for the test subject groups; only the framing changed. That let the researchers attempt to isolate the psychological effect of calling the system a coworker like “Alex.”From the study: "The main task involved reviewing five documents with errors, reviewing job descriptions for managers in HR and reviewing budget documents for managers in Finance. Regardless of domain, all
participants reviewed a sequence of five documents using an interface that allowed for highlighting, flagging, and commenting. Participants were given a total of 20 minutes to review as many
documents as possible."
 |
| Clker.com |
A few things of note.
- Exactly who the subjects were: Using "managers and executives" as subjects, instead to those "one level down" who probably have much more experience creating such documents using AI. There is no real demographic or qualification-based information on them presented.
- The vague description given of the nature of the model documents involved, how they were created and the types of errors encoded, critical to understanding the main effect here. (The actual data and protocols from the research are not publicly available that I can determine.)
So, given that . . . The results were still striking at face value:
- ALEX-3's work got 18% less errors corrected by those who already work in a system that has an AI-bot as a functional member of the team, about a third of the subjects (~400)
- Almost half of the subjects said they'd probably "send the piece on or "up" for someone else to deal with.. .
From the study conclusion: "Among managers whose organizations have AI agents on their org chart, describing identical drafts as coming from an AI employee reduces error detection, increases reliance on additional review, and shifts perceived accountability away from the manager and toward the AI system."
Wiles et al speculate that the effect may be coming, in part, from "anthropomorphisizing" of the bot (giving it some human-sounding agency and identity)--which could mean many things, of course. It could also be a reflection of the impact on the entire system of having AI assuming various functions, where there may be different or additional individuals responsible for engaging with or mediating the AI output.
They go on to recommend not doing that, not calling AI a team member . . . That seems simple enough.
Just got a great comment from a friend, Kensaku Yoshida: "Thanks for the interesting article. It seems very likely that the more we use AI for a specific task, the more the AI will accumulate information about how we perform the task, (italics, mine) thus gradually making it ever more difficult for us to distinguish its behavior from our own. I’m reminded of Kazuo Ishiguro’s famous novel ‘Klara and the Sun.’
Summary/critique
It is safe to assume that those managers and execs approached ALEX-3's output through a well conditioned set of filters that tend to trigger at least some diffusion of authority, let alone perhaps mild panic for the less competent writers and "editors" themselves. On that, I do concur with the researchers, but research on just the perceptual and cognitive impact of AI-generated text or communication of almost any kind, has clearly established that the human brain is not all that well set up to handle both the textual and relational dimensions involved in a task like this (as observed by Yoshida.)
Just the nature of the task, confronting AI text output can be daunting, especially if you do not have sufficient background in the topic. All the texts were hypothetical pieces, not organic to the work of the reviewer, although the form and purpose of the texts would, we assume, have been very familiar to them. Since the text was NOT AI-generated in the first place, there could have been some additional "indicision" related to critiquing it as such. The "seriousness" and location of each error may figure in. The texts could well have been constructed explicitly to evoke the main effect, devised to be relatively ambiguous as to human or AI creation. (As a linguist that issue looms large for me. The protocols and data from the study are not accessible to the public as far as I can tell.)
But this is still a useful piece. The integration of AI and the controlling entities behind it into all aspects of our experience is happening far more quickly than all but a few of its "creators" ever imagined.
There are an almost infinite number of other recommendations on how to cope with AI today, including my own AI-ROBICs project, in addition to just not calling bots, buddies . . . and to quote Yoda (of Star Wars): "There is only DO or do NOT. There is no try."
-------------------------------------------------------------------
Caveat emptor! I always conclude my reports, as you know, with recommendations, generally that you seriously consider signing up for Legalshield (best choice or one of its competitors) and effective web traffic protection, such as IDshield.
Things are changing.
IDshield is still excellent, especially at taking care of what you could possibly piece together for yourself if you had time and web savy. Legalshield or a good lawyer in the family or on (affordable) retainer, on the other hand, is becoming critical in not just handling typical, everyday legal issues, but in being capable of responding to the sophisticated level of attacks and scams that AI now facilitates.
Along, of course, in addition to following this blog for the latest in personal AI "counter attack" ideas!
Do check out these new introductory videos from Legalshield about the standard subscription system and what the company is about. Then get in touch and I'll show you how to join up, too, or pick up a new side gig that can easily develop into a great small business opportunity.
wracton@gmail.com
williamacton.legalshieldassociate.com
No comments:
Post a Comment