December 4 News.AmazonCloudTech (AWS) has released a new tool designed to address the problem of hallucinations generated by AI models.
1AI notes that at re:Invent 2024 in Las Vegas, AWS introduced the Automated Reasoning checks tool.The tool validates the model's response accuracy by cross-referencing information provided by the client.
AWS claims that this is the "first" and "only" protection against hallucinations. However, this claim may not be accurate.Microsoft's "Corrections" feature, introduced this summer, is almost identical to "Automated Reasoning Checks"., both of which can flag AI-generated text that may have factual errors. Google's Vertex AI platform also offers a tool that allows customers to make modeled responses more reliable by using data from third-party providers, their own datasets, or Google searches.
"Automated Inference Checking, available through AWS's Bedrock model hosting service (specifically the Guardrails tool), tries to figure out how the model arrived at an answer and determine if the answer is correct. Customers upload information to establish a fact base, and then Automated Reasoning Check creates rules that can be optimized and applied to the model.
When the model generates answers, Automated Reasoning Check validates them and uses the factual basis to arrive at the correct answer in the event of a possible hallucination. It presents this answer along with possible incorrect answers so that the client can see how far the model deviates from the correct answer.
AWS says PwC has begun using "automated inference checking" to design AI assistants for its clients, and Swami Sivasubramanian, vice president of AI and data at AWS, hints that this type of tooling is what attracts clients to Bedrock.
But one expert said this summer that trying to eliminate the illusion of generative AI is like trying to eliminate hydrogen from water, Techcrunch reports.AI models create illusions because they don't actually "know" anything. They are statistical systems that recognize patterns in a series of data and predict what the next data will be based on previously seen examples. Thus, the model's response is not an answer, but a prediction of how the question should be answered - within a certain margin of error.
AWS claims that Automated Reasoning Check uses "logical accuracy" and "verifiable reasoning" to reach its conclusions, but the company has not provided any data to demonstrate the reliability of the tool. .