An End To Black-Box AI Hallucinations?
AI will quickly become a dystopian nightmare, unless it is based on tamper-proof blockchain ledger
AI could prove to be detrimental to society if it is not transparent, verifiable, and decentralized. Black-box AI hides its biases, proliferates deepfakes and misinformation, and introduces massive single points of failures. The only way to reliably achieve a verifiable is with blockchain technology.
Blockchain can mitigate AI hallucinations, where models create false, made-up or illogical information while framing it as a fact. It can also reduce biases, and unbridled creation of content. It is verification capacity, not creation capacity, that will prove to be the limiting factor for AI.
Language models hallucinate due to training and evaluation procedures still rewarding guesses over recognizing uncertainty. It’s possible that even with error-free training data a language model could still produce errors or include useless statements.
And in scaled-up AI models, the problem becomes worse. The rate of imitative falsehoods increases. Not to mention, LLMs have a proclivity to produce more hallucinations as it tries to maintain coherence with previous hallucinations. And, in the more complex system, the errors are more difficult to recognize.
LLMs refined by humans also produce outputs that differ from their implicit designs, called sycophancy, in which it produces answers of which it thinks a user would approve. This undermines truthfulness in the outputs.
In order to work reliably, AI requires meaningful human review of its outputs. At present, AI algorithms are opaque. It is unclear why particular determinations are made, which undermines trust, especially in critical applications like self-driving cars, health insurance decisions, high speed frequency trading, and national security.
Traditional crowdsourcing suffers from the same lack of transparency due to its centralized, proprietary platforms that conceal how tasks are allocated, workers are compensated, and metrics are evaluated.
This naturally leads to unfairness in assigning tasks and rewarding workers. Additionally, the quality of labeled training data cannot be guaranteed. An algorithm that regulates inputs must accurately calculate the annotations between multiple independent inputs, in order to guarantee fairness, quality, and verifiable rewards.
Alternatively, blockchain maintains an immutable ledger of who made the call, who checked it, and what the outcome was, while not storing all of the raw data in a honeypot that can be targeted by hackers with automated bots, script kiddies, state actors, and organized crime. Blockchain proffers an open, decentralized ecosystem with transparent, tamper-proof records of linguistic-based rules and collaborative human-like verification, rather than black box probabilistic models.
Data models would be recorded onchain for the sake of accuracy and immutability of verification as a means of minimizing trust issues.
A blockchain system executes code automatically and instantaneously, publicly and traceably. It combines privacy-preserved truth discovery from the workers tasked with labeling the data.
Blockchain counterbalances AI’s centralization through decentralized data management, computation, and governance. This makes for increased inclusivity, transparency, and user privacy with verifiable audits of data inputs and model outputs.
Blockchain offers full contribution traceability, usage data, compute, models, and inferences, all while not exposing sensitive data. AIs trained on blockchain-verified data have already shown a 40 percent reduction in biased outcomes, with no data manipulation detected.
At the heart of AI is an ongoing dynamic feedback loop between humans and AI systems, which gives rise to a co-evolution, rather than an AI takeover. Each influences the other with far-reaching social implications. AI systems must foster fairness, inclusivity and societal well-being.
Without transparency, verifiability, and decentralization of blockchain, AI will continue to produce unreliable outputs. Only blockchain fixes that.
