AI models struggle with basic physical tasks in manufacturing

Frontier AI models continue to struggle with physical tasks despite significant advancements in other domains, revealing a persistent gap between AI’s digital capabilities and real-world applications. A recent evaluation of leading AI models attempting to create machining plans for brass parts demonstrates that even the most advanced systems fail at basic manufacturing tasks that human machinists routinely complete. This disconnect highlights a potential future where white-collar knowledge work becomes increasingly automated while physical jobs remain protected from AI disruption.

The big picture: Frontier AI models demonstrate alarming deficiencies in both visual perception and physical reasoning when asked to create machining plans for a basic manufacturing task.

• Most tested models failed to correctly identify obvious features like holes and flats in the part design, with only Gemini 2.5 Pro showing limited progress by identifying major features in about 25% of attempts.
• Even when models could “see” the part correctly, they proposed physically impossible manufacturing sequences that would damage equipment or produce unusable parts.

Why this matters: These limitations suggest an emerging “automation gap” where knowledge work becomes increasingly automated while physical and manufacturing jobs remain largely AI-resistant.

• This divide could create significant economic and social tensions as different sectors experience uneven automation impacts.
• Nations with strong industrial bases may gain geopolitical advantages as physical skills retain their value against AI disruption.

Behind the failures: AI’s struggle with physical tasks stems from their lack of tacit knowledge that human machinists acquire through hands-on experience.

• Models failed to account for real-world considerations like metal rigidity, machine vibration (chatter), and the physical constraints of workholding during machining operations.
• AI systems regularly suggested physically impossible sequences of clamping and rotation, demonstrating fundamental gaps in spatial reasoning.

The improvement challenge: Training AI on physical tasks presents unique obstacles compared to purely digital domains.

• Physical tasks lack clear reward signals that can be easily optimized through traditional AI training methods.
• Learning through trial-and-error in manufacturing environments would be prohibitively expensive and potentially dangerous.
• The feedback loop between action and result in physical environments is complex and difficult to model.

Between the lines: The persistence of these limitations suggests that the future of work may not follow the commonly predicted pattern of universal AI displacement.

• While digital and remote knowledge work faces accelerating automation, physical and manufacturing jobs may remain viable human careers for much longer than anticipated.
• This challenges the prevailing narrative that all types of work will be equally vulnerable to AI disruption.