Honest Lying: Why Scaling Generative AI Responsibly is Not a Technology Dilemma in as Much as a People Problem

In the 24/7 news cycles of generative AI hype and fears of widespread job loss, there’s no attention being paid to herculean hiring effort and human expertise necessary to scale generative ai responsibly.

SwissCognitive Guest Blogger: Caryn Lusinchi, AI Strategy & Governance Advisor, Founder of Biasinai.com – “Honest Lying: Why scaling generative AI responsibly is not a technology dilemma in as much as a people problem” Image source: Jametlene Reskp on Unsplash

With the ultimate widespread adoption of LLM, GAN and multi-modal generative AI over the next few years, all AI products and services will have warning labels attached:

WARNING: We may not produce output that is real. Believe at your own risk. You revoke your right to sue us if you reproduce conversations, images and videos for high risk use cases that cause hate, harassment, violence, self-harm, illegal activity, deception, discrimination, spam, etc.

Given the recent Google Bard James Webb telescope faux pas combined with #chatgpt ’s public beta that confidently answered with facts that were not historically true, not scientifically based or not physically possible on earth, the majority are categorizing this behavior as “hallucinating”.

Confabulating

However, others have alluded to ‘confabulating’, a more specific term that is typically associated with Alzheimer’s disease and dementia.

ClinMedjournal defines it as “the production or creation of false or erroneous memories without the intent to deceive, sometimes called “honest lying. Alternatively, confabulation is a falsification of memory by a person who believes he or she is genuinely communicating truthful memories.” WebMD adds, “Confabulation is caused by brain damage or poor brain function, but researchers are unsure which parts of the brain are at fault.”

It is only intuitive that we as a species, anthropomorphize AI (when we attribute human characteristics or behavior to objects), but we often forget that we are humans who are still interfacing with machines. Machines that are only as reliable as their millions of parts will inevitably break down and need repair. AI doesn’t work like a human brain does and only has limited memory in that it stores past data and predictions to improve future predictions.

One common example of confabulation is Dall-E’s image output to a text prompt, “a salmon swimming down a river”

The AI can’t understand symbols without context. It produced exactly what you requested. Salmon swimming down a river.

It was never trained on a dataset of images to understand real-world models. It never learned the complex ecosystem of meaning. It doesn’t know the difference between live salmon spawning upstream in ice-cold Alaskan Rivers vs. gutted salmon filets sold by weight in the seafood section of your local supermarket. Now think of the millions and billions of data points in a generative AI database that have absolutely no context.

Fish are an innocent example, but imagine if an individual is subject to a generative AI outcome that influences a high-risk use case within healthcare/medicine, employment decision-making, law enforcement, government or financial services industries..

In a March 2022 Future of Life podcast with Anthropic’s Daniela Amodei, she noted,

”Some of the AI safety research we do is that different sized models exhibit different safety issues…. Some of the larger models that we’ve seen, we basically think that they have figured out how to lie unintentionally. If you pose the same question to them differently, eventually, you can get the lie pinned down, but they won’t in other contexts… there are quite a lot of behaviors emerging in generative models that I think have the potential to be fairly alarming.”

Demand will outstrip supply for a decade, maybe more

In the 24/7 news cycles of generative AI hype and fears of widespread job loss, there’s no attention being paid to herculean hiring effort and human expertise necessary to scale generative AI responsibly.

Enterprises will need to hire data science/ML engineering experts to continually train, test, launch and retrain models coupled with the thousands of global in-house or 3rd party Mechanical Turk contractors required to red team models, screen/moderate output and escalate problematic content.

And right now, there’s a dearth of Data Scientists and Machine Learning experts globally. At Gartner’s Data & Analytics conference in August 2022, the word talent “shortage” was replaced by “crisis”.

As of February 2023, LinkedIn has 884,810 open global roles requiring data scientist skills and 202,856 results for machine learning engineers.

That’s currently 1 million worldwide jobs where demand outstrips supply and the gap isn’t closing anytime soon. According to the US Bureau of Labor Statistics, data scientists are currently in the top 20 fastest-growing occupations in the U.S. with 31% projected growth over the next 10 years, and about 11.5 million new jobs in the field by 2026.

Given the average tenure of a data scientist at a company is 1.8 years, startups and enterprises are additionally facing a future of huge technical debt and retention issues until which time STEM and higher education can re-skill students in DSML/AI fields.

The August 2022 “Preventing an AI catastrophe” article by Benjamin Hilton on the 80,000 hours blog points out that technical AI safety and AI strategy/policy research and implementation are vastly neglected fields.

“We estimate there are around 400 people around the world working directly on reducing the chances of an AI-related existential catastrophe (with a 90% confidence interval ranging between 200 and 1,000). Of these, about three quarters are working on technical AI safety research, with the rest split between strategy (and other governance) research and advocacy.”

Will there be enough experts to enforce AI regulation

The EU AI Act, Brazil’s Draft AI Law and other evolving worldwide regulations require the very same subject matter experts that enterprises are currently struggling to recruit. This is compounded by the reality that Google and Microsoft are seducing DSML engineering talent away from both government and enterprise with competitive offers.

There has been a global rally cry for AI regulation. However, when it does pass and become standard, statistics show there will be a shortage of independent subject matter experts who understand the complexity of the technology to investigate it and adequately enforce it. Resources will be pooled to police only the most egregious, front page headline cases.

Non-profit organizations such as ForHumanity.center anticipate the AI market will grow creating demand for certified auditors so they have recently launched accreditation courses around jurisdictional data & AI law as well as risk frameworks. They believe the governance of AI and autonomous systems should be audited with independent and crowdsourced criteria.

On the other hand, there are a handful of smaller boutique consultancies such as BABL.AI, ORCAA and BHH.AI conducting 3rd party independent audit services and increasingly, for-profit consultancies such as Deloitte, Accenture and McKinsey are showing signals on job boards that they’re staffing up AI auditing branches.

Multiple choice answers to the people problem

Software vendors like Databricks, MonteCarlo, Weights and Biases Truera and Robust Intelligence, are building the next generation of AI data testing, validation and monitoring tools for the next generation of data science/ML engineers. Anticipating the lack of technical experts, they’re making the workflow process simpler by developing user friendly templates and low code/no code features to appeal to ‘data citizen’ employees. Data citizens may come from traditional data & analytics or DevOps backgrounds. They may have less technical hands-on experience but curiosity to pivot careers.

Second, enterprises are launching In-house DSML accelerator programs that financially incentivize current employees and offer access to education, resources and mentorship to grow and nurture their talent pipeline internally.

Third, universities are expanding Master of Science programs to include Artificial Intelligence and multi-institution research groups are creating niche offerings such as Stanford’s Center for AI Safety, University of Oxford’s Future of Humanity Institute and the Center for Human Compatible Artificial Intelligence.

Finally, academia-enterprise and academia-government partnerships are helping identify junior level talent for workforce internships and establishing agreements to pilot PHD-level R&D lab projects for tactical real-world application.

In the generative AI hype we’ve forgotten, ourselves.

The workforce transformative powers of generative AI is everywhere. How emails, pitch decks, school essays, content marketing, design, music composition, movie production will revolutionize human productivity and efficiency. And yes, inside Reddit, Discord discussions, among VCs and at industry conferences, the top common scalability challenges for generative AI have been acknowledged: computing power, data privacy-security, data scarcity, bias and more.

However, in our laser focus on magical technology itself, we’ve overlooked the massive recruiting challenges required for it to scale responsibly.

Us, the humans.

About the Author:

Caryn Lusinchi is FHCA (For Humanity Certified Auditor) under UK, EU GDPR and NYC Bias Law and has a certificate in Foundations of Independent Audit of AI Systems (FIAAIS). She is the founder of Biasinai.com.