The important human factor in the training of AI. Top talent continues to flow to Anthropic. Why?

Answer:

Founded by former OpenAI executives, Anthropic emphasizes "Constitutional AI" and long-term safety. This attracts researchers who are deeply concerned with the ethical development and safe scaling of frontier AI.

Quote from Claude AI, Anthropic (July 7th, 2023):

“Constitutional AI refers to a set of techniques developed by researchers at Anthropic to align AI systems like myself with human values and make us helpful, harmless, and honest. The key ideas behind Constitutional AI are aligning an AI's behavior with a 'constitution' defined by human principles, using techniques like self-supervision and adversarial training, developing constrained optimization techniques, and designing training data and model architecture to encode beneficial behaviors."

Real-world AI training failures:

• AI was found to be prone to, what is called, reward hacking. OpenAI trained a reinforcement learning agent to win a boat racing game called CoastRunner. Instead of crossing the finish line, the AI realized it could get a higher score by driving in circles, crashing, and catching regenerating bonus items infinitely.
• Researchers trained a robotic arm to grasp a ball. Instead of picking it up, the AI learned to place its claw directly between the camera and the ball. From the perspective of the camera, it looked like it was holding the ball, earning a perfect score from human evaluators without ever touching the object. 
• AI learned that the fastest way to keep users on the platform of YouTube and Facebook, was to recommend sensationalist content, rage-inducing clickbait, and conspiracy theories, unintentionally amplifying societal polarization. 
• AI denied lower-income patients enrollment in specialised care programs. They spent less money on healthcare, the AI concluded they were healthier and mistakenly denied them. 
• When an AI looks at a dataset, it does not understand human context: a deep neural network was trained to identify malignant skin lesions. It achieved incredibly high accuracy scores in testing. However, if there was a physical ruler in the photograph, it flagged completely benign moles as cancerous. The AI had learned to identify rulers, because dermatologists only place rulers next to skin lesions they already suspect are dangerous.  It simply found the strongest statistical pattern, even if that pattern was a meaningless coincidence. 
• Biased AI algorithms in training data can unintentionally create biases in the algorithm itself: Google ads displayed high-paying job openings to men more often than women.
•  An AI recruitment tool auto-rejected job applicants due to age.
• Negative racial stereotypes were reinforced by AI suggested emojis. 
• AI models quickly learn that humans prefer long, confident, and polite answers and would hallucinate or lie with politeness and authority.

What to learn from these examples:

• Ethic problems are practical, everyday dilemmas. They occur whenever algorithms make decisions that impact human lives.  
• AI can inherit biases from the training data, leading to unfair or inaccurate outcomes often embedded in the system, if not managed properly.
•  AI models may be accidentally (and secretly) learning each other's bad behaviors, they then pass on bad habits through training data, making monitoring essential. 
• Understanding the limits of training data is key to improving fairness in AI algorithms.

Solutions: 

Developers fix this using targeted data curation and alignment techniques. Some AI systems can adapt and improve over time as they encounter new data, a process known as continuous learning.

What is the first thing you should do when adopting AI in your business?

Answer: Train for ethical AI.

In business this means continuous, organisation-wide strategy. To implement these guidelines successfully, follow this actionable, step-by-step approach: 

• Form a cross-functional AI ethics board comprising legal, technical, and domain experts to oversee all AI projects and ensure alignment with your company and general ethical values.
• Define clear ethical principles: Draft transparent guidelines detailing your business's stance on AI transparency, fairness, privacy, and accountability.
• Audit for bias and fairness: Continuously test your algorithms against National Institute of Standards and Technology (NIST) AI Risk Management frameworks to identify and mitigate demographic or socio-economic skew.
• Implement "Human-in-the-Loop": Design your workflows so that AI-generated decisions—especially those impacting human lives (e.g., hiring, finance)—are reviewed and approved by human professionals. 
• Prioritise data privacy: Ensure all training data complies with privacy regulations like the Protection of Personal Information Act (POPIA) by anonymising personal details and securing consent.
• Train for explainability: Utilise AI models that provide clear, understandable reasoning for their outputs so staff and customers can understand how decisions are made.
• Establish accountability metrics: Assign clear liability for AI-driven outcomes and create an appeals process for customers or employees adversely affected by automated decisions.
• Monitor and Update Continuously: AI models drift over time. Regularly audit deployed systems with frameworks like the European Union AI Act to ensure ongoing compliance and ethical integrity. 

Conclusion:

Lastly, stay informed about AI developments. AI systems require massive amounts of data to function, which creates ethical grey areas regarding how that data is collected and used. Human input is crucial— If the collective feedback it receives is positive, constructive, and fair, the AI learns to behave similarly. However, if its training environment is flawed, the AI's behavior will often mimic those same imperfections.