Apple Reveals Secret Strategy To Handle AI Hallucination And Build More Human Conversations

Apple introduced Massive Multitask Agent Understanding (MMAU), (photo: dock.apple)

JAKARTA Amid the dominance of OpenAI, Google, and Meta in the artificial intelligence race (AI), Apple is quietly making big breakthroughs that can change the game. Through a series of recent scientific research, Apple researchers announced eight important papers focused on understanding AI errors, deepening personalization, and ways to reduce the hallucinations of large language models (LLM).

This move is a strong signal that Apple not only wants to be a passive player in the AI world, but also wants to create a more ethical, humane, and reliable AI ecosystem and of course be closely integrated with Apple products in the future.

In one of its latest papers, Apple introduced the Massive Multitask Agent Understanding (MMAU), a comprehensive evaluation system for measuring LLM capabilities in five important areas: understanding, reasoning, planning, problem solving, and self-correcting capabilities.

The MMAU includes 20 tasks consisting of more than 3,000 unique prompts providing clearer and more standardized benchmarks, compared to previous approaches that Apple said were still confusing and insufficient to identify AI's root of error.

The goal is to improve AI capabilities by understanding where the error came from, Apple researchers wrote in a paper published in Cornell's scientific archives.

So far, AI has difficulty understanding users' conversations personally due to limited memory of long-term conversations. Answering this challenge, Apple developed a system called Pipeline for Learning User Conversations in Large Language Models (PLUM).

Instead of simply remembering the user's small preferences, PLUM extracts questions and answers from previous interactions and injects them into the system, creating AI that feels closer and relevant to users personally.

This technology opens the door for Apple's AI to provide conversational experiences that are not only sophisticated, but also feel humane and consistent.

Apple Uses External Validation

Another main problem with AI is how models can convey answers as if they are absolutely true, even though they often contain errors or biases.

In a paper titled Can External Validation Tools Improve Annotation Quality for LLM-as-a-Judge?, Apple proposes the use of external verification tools, such as web searches and code executions, to re-examine the correctness of AI answers.

Although the results are not 100% accurate, this approach has proven to often improve the quality of answers and could be an important foundation for more transparent and trustworthy AI.

Apple not only published the paper, but also released eight video presentations from an internal workshop entitled Human-Centred Machine Learning 2024. The videos review AI interfaces, UI understanding, and AI personalization in the context of daily use.

More than that, Apple will appear at the Association for Computational Linguistics (ACL) annual conference in Vienna on 27 July 1 August 2025. There, Apple will present and sponsor 18 workshops that carry the theme of a more humane and responsible AI.

So far, the public thinks Apple is lagging behind in the AI race because there is less attention than competitors. However, this series of papers and innovations proves that Apple is actually building AI in depth, be careful, and side with users.

With an approach that focuses on ethics, reliability and personalization, Apple seems to be preparing for an AI breakthrough that will not only compete in the market, but could reshape industry standards.

"Apple wants a more trustworthy AI, better understands humans, and is free from fatal mistakes. Not only smart, but also wise," said Jeffrey P. Bigham, Director of Human-Centred Machine Intelligence at Apple.

With all this, the question now is no longer 'Is Apple ready to enter AI?' But more precisely: 'Is the world ready to welcome Apple's version of AI?'