065 - Cognitive Bias Detection Study

After 2.5 months the Cognitive Bias Detection Study paper covering phases 1 and 2 is now listed on the SSRN pre-print for Heliyon/Cell Press/Elsevier. The phrase "Moving at the speed of peer review" comes to mind.

For those who missed it previously, the first phase of the study established a baseline of human performance at cognitive bias detection, comparing it to that of the detection system, and creating a collective human detection for further validation. The key takeaway from this phase was that the system performed better than the average human when considering the full spectrum of cognitive biases, and nearly as well as the very best human when performing detection on the 75% of categories where performance was strongest.

Phase 2 of the study applied the detection system to the outputs of the top 5 LLMs at the time, including GPT-4, Bard (PaLM2), Claude, Falcon, and Vicuna. Relative to the patterns observed in detections from human-generated text, these LLMs showed notably higher levels of bias in roughly half of all categories and only consistently displayed lower levels of bias in one.

Next, Phase 3 will focus on detecting patterns of cognitive bias expression over time from specific individuals. Over 5,100 blocks of text, each one or two sentences in length, have been gathered from 50 individuals in 4 categories for this purpose, extracted from interviews with each individual. The first group includes known and convicted frauds like Bernie Madoff. The second includes several rather bloody individuals like Stalin. A third "Test" group includes a few noteworthy CEOs, authors, and other influencers. The final group is a collection of respected scientists including a few Nobel Prize winners.

This line of research will hold a deep significance for the entire sum of human decision-making, though most likely won't comprehend as much for some time yet.

Humans routinely attempt to have 15 to 30-minute conversations to determine if an individual or group is competent, heavily biased/delusional, or fraudulent, but they most often have very little success in this regard. HR hiring processes, investor pitch meetings, and meetings with potential clients are all examples of this.

Though it remains to be seen if Phase 3 will produce the specific means of accomplishing this, nothing is preventing this goal from being accomplished with greater accuracy and robustness than humans demonstrate in pursuit of it today.

Humans have grown adept at fooling one another in this regard, as they've optimized themselves to manipulate other humans. Consequently, that optimization makes them both more visible and more vulnerable to any other robust perspective.

The as-of-yet unconvicted Bernie Madoffs of today's AI industry are operating on borrowed time.