DeepSeek
DeepSeek
Extremely cheap reasoning at scale.
Impact Indicators
Job Displacement
No verified data attributes job losses to DeepSeek. Available labor research measures occupational exposure and automation potential across generative AI broadly, not employment outcomes caused by DeepSeek or any single model.
Privacy & Surveillance
Italy’s data protection authority restricted processing of Italian users’ personal data by DeepSeek’s operating companies after raising concerns about data collection, legal basis, retention and storage. Government device restrictions elsewhere are policy actions, not findings of privacy violations.
Water Usage
No measured DeepSeek-specific water-use figure exists. Research documents facility and regional water demand from Chinese data centers and models water use for AI workloads more broadly, but those figures cannot be attributed directly to DeepSeek.
Energy & Carbon
No independently verified energy or emissions figure exists for DeepSeek’s production infrastructure. Peer-reviewed testing of locally run DeepSeek R1 variants found that reasoning-heavy responses can require substantially more energy than concise responses, but those laboratory results do not measure DeepSeek-V4 or the company’s production fleet.
U.S. reporting requirements do not yet provide standardized, model-level disclosure for many AI human and environmental impacts. When U.S. evidence is incomplete, DoctorKnow uses verifiable EU regulatory disclosures, public-sector data, independent research, and reputable journalism to provide the strongest evidence currently available.
Evidence scope is identified as model-specific, provider-level, facility-level, regional/grid, or industry-level.
Scores use a 0–100 scale. Higher scores indicate stronger performance or lower concern; 100 is the best possible rating.
Reasoning
91
Coding
92
Writing
82
Analysis
88
Speed
84
Scores use a 0–100 scale, where 100 represents the strongest outcome and 0 the weakest.
What DeepSeek does
DeepSeek is built by DeepSeek and is positioned for code, analysis and research. Our readings score it strongest on code generation and multi-step reasoning.
Its weakest measured axis is long-form writing (82/100), so plan around that when the work depends on it. Access is open, and the weights are published, so the model can be inspected and self-hosted.
Ethical stance
Transparency reads 58/100. Some model cards and evaluations are public, but training data and methods are only partly described.
Labor and job-displacement posture reads 34/100. Data-annotation conditions and displacement impact are largely unaddressed.
Regulatory posture reads 26/100. The vendor's public posture leans against binding oversight.
Privacy and data handling
Privacy and data handling reads 38/100. Prompts may be retained and used for training by default, with limited user control.
Because the weights are published, the strongest privacy option is self-hosting: prompts never leave infrastructure you control.
Energy, carbon and water
Carbon footprint reads 44/100. The energy mix is mixed, and disclosure covers some but not all serving regions.
Water and local impact reads 39/100. Cooling draws meaningful water in drought-exposed regions, or usage is not reported.
Speed and efficiency also matter here: a 84/100 speed reading means fewer compute-seconds per answer for routine work.
Diagnostic notes
Open weights, but opaque data provenance and no environmental disclosure.
Readings are DoctorKnow.ai's own 0–100 assessments, compiled from vendor documentation and public reporting, and revised as new disclosures land. They are a comparison aid, not a certification.