Way back in 2017 Andrew Ng proposed a “highly imperfect rule of thumb”, that “almost anything a typical human can do with less than one second of mental thought, we can probably now or in the near future automate using AI“.
In AI world 2017 is light years ago, but if we add a bit to the sentence above it is still a good mental model for understanding what AI can do: Anything repetitive task a typical highly trained human can do with less than one second of mental thought without extraordinary effort, we can probably now or in the near future automate using AI*.
*terms and conditions may apply, does not cover empathetic tasks, advanced reasoning, one-shot learning, perfect recall/precision, and may cause human harm if not applied ethically.
To quantify this statement further here are some reference points for where AI systems shine and where they are mostly useless (today and likely in the near future).
Standard use-cases
Visual Processing
- Defect spotting in manufacturing (surface scratches, misalignments, etc.)
- Object identification in images (people, animals, vehicles, etc.)
- Face recognition and verification
- Reading text from images (OCR)
- Classifying image content (indoor/outdoor, day/night, etc.)
- Detecting emotions from facial expressions
- Identifying brand logos or specific objects
- Screening medical images for anomalies (X-rays, CT scans)
Language Processing
- Translation between languages
- Transcribing speech to text
- Summarizing text content
- Sentiment analysis of text
- Categorizing documents by topic
- Grammar and spelling correction
- Answering factual questions
- Language identification
Audio Processing
- Voice recognition/speaker identification
- Music genre classification
- Detecting specific sounds (glass breaking, gunshots, etc.)
- Speech emotion recognition
- Background noise classification
- Audio quality assessment
Data Analysis
- Anomaly detection in time series data
- Credit card fraud detection
- Customer churn prediction
- Basic pattern recognition in structured data
- Simple classification tasks (spam/not spam)
- Recommendation systems based on past behavior
Unexpected use-cases
Hidden Patterns That Humans Miss
- Predicting equipment failure before visible signs appear (by detecting subtle vibration patterns)
- Identifying early-stage diseases from biomarkers that doctors might overlook
- Detecting fraudulent financial transactions using patterns across thousands of variables
- Identifying authorship based on subtle writing style fingerprints
- Predicting consumer behavior shifts before they appear in sales data
Time-Compressed Tasks
- Analyzing thousands of hours of surveillance footage in minutes
- Processing millions of research papers to find connections humans haven’t made
- Reviewing large legal contracts faster than specialized lawyers
- Scanning decades of climate data to identify subtle trends
- Analyzing satellite imagery over time to detect small environmental changes
Counter-Intuitive Correlations
- Predicting successful movies based on seemingly unrelated script elements
- Identifying at-risk students using behavioral patterns invisible to teachers
- Determining optimal pricing strategies from hundreds of market variables
- Finding unexpected drug interactions across diverse patient populations
- Predicting material properties without running physical experiments
Creative Domains
- Generating novel molecular structures for drug discovery
- Creating artwork in specific styles that fools art experts
- Composing music that resonates emotionally with listeners
- Writing code that outperforms human programmers in specific tasks
- Designing optimized physical structures humans wouldn’t conceive
Complex Integration Tasks
- Understanding context across multiple data types (text, image, numerical)
- Translating between specialized domain languages (technical, medical, legal)
- Identifying personality traits from combinations of behaviors
- Detecting subtle audio/visual mismatches in deepfakes
- Combining multiple sensory inputs to navigate complex environments
Where AI (mostly) fails
Common Sense Reasoning*
- Understanding why you can’t put an elephant in a refrigerator
- Temporal reasoning: “I dropped the glass, and it shattered on the floor, Was the glass shattered before it hit the floor?”
- Object permanence: “If you put a toy under a box and walk away, is the toy still there?”
- Understanding basic physics of everyday objects
*Newer models and the introduction of self-thought is slowly encroaching on this area. Current models have some success in common sense reasoning, but we also experience spectacular failures. Proceed with caution.
Contextual Understanding
- Detecting sarcasm or irony consistently
- Understanding cultural references without explicit explanation
Grasping metaphors in context (“that project is a sinking ship”)- Following conversations with implied information
Spatial Reasoning*
- Understanding how to pack items efficiently in a confined space
- Navigating through cluttered environments without collisions
- Predicting how objects will interact when manipulated
- Understanding which objects can fit through which openings
*Again multimodal LLMs with image capabilities are starting to have some mimicking of spatial understanding, but there is currently no understanding of perspective and occlusion.
Adapting to Novel Situations
- Generalizing knowledge to entirely new contexts
- Solving problems using tools in creative ways
- Adapting to unexpected changes in task parameters
- Handling exceptions to rules without explicit programming
Physical Tasks
- Manipulating unfamiliar objects with appropriate force
- Tying knots or handling flexible materials
- Working with transparent or reflective objects
- Performing fine motor skills in varying conditions
Ethical and Social Judgments
- Understanding appropriate behavior in different social contexts
- Recognizing when to make exceptions to rules for ethical reasons
- Balancing competing values in complex situations
- Judging when humor is appropriate vs. inappropriate
Intentionality and Goals
- Understanding human motivations behind actions
- Distinguishing intentional from accidental behaviors
- Recognizing deception or manipulation
- Inferring long-term goals from short-term actions
What does it all mean?
This means that despite the major progress we have seen in AI over the last couple of years AI is still mostly a productivity booster. It needs a human in the loop for most things to create a meaningful impact. It needs specific constraints and clear interpretation. It is excellent as a narrow intelligence, it is bad as a strategic thinker.
The most effective implementations combine AI’s computational power with human judgment, creativity, and ethical reasoning. We’re seeing this hybrid approach succeed across industries—from healthcare diagnostics augmented by radiologists to legal document analysis guided by attorneys.
The strategic question isn’t whether AI will replace humans, but how we can design systems that maximize the unique strengths of both human and artificial intelligence.
I am excited to see how we may need to shift Andrew Ng’s statement in 2030, maybe it will sound something like this: “Any task done by AI can match or exceed human capability in execution, but requires human oversight proportional to its complexity and potential impact.”