Legal documents represent the highest-stakes category for translation. A contract clause mistranslated could void an agreement, shift liability, or create obligations never intended. When law firms and corporations consider machine translation for legal work, they are gambling that accuracy is high enough to trust.
DeepL markets itself as the most accurate machine translation service. For general content, this claim holds well. For legal content, we wanted rigorous evidence, not marketing claims.
We tested DeepL on 500 legal contracts across multiple jurisdictions, document types, and language pairs. Here is what accuracy looks like when the stakes are highest.
Key Takeaways
- Overall accuracy on legal contracts was 87.6%, significantly lower than general business content
- Negation errors (turning prohibitions into permissions) appeared in 12.4% of contracts
- Jurisdiction-specific legal concepts showed the highest error rates
- Machine translation alone is not safe for any legally binding document
- A structured human review workflow can make AI-assisted translation viable for certain legal tasks
Testing Scope and Methodology
What We Tested
We assembled a dataset of 500 legal contracts from multiple sources:
Service agreements (150 contracts): Vendor contracts, service level agreements, and professional services agreements in English, German, French, and Spanish.
Employment contracts (100 contracts): Offer letters, employment agreements, and severance agreements in English, German, French, and Japanese.
Non-disclosure agreements (100 contracts): Mutual NDAs, one-way NDAs, and multi-party NDAs in English, German, French, Spanish, and Portuguese.
Partnership agreements (75 contracts): Joint venture agreements, strategic partnerships, and co-development contracts in English, German, and French.
Regulatory documents (75 contracts): Compliance attestations, regulatory filings, and government contract provisions in English and German.
Evaluation Process
Two qualified legal translators with experience in commercial law evaluated each translation. They worked independently, then reconciled differences through discussion.
The evaluation focused on:
Meaning preservation: Does the translation convey the same legal meaning as the source?
Terminology accuracy: Are legal terms translated correctly for the target jurisdiction?
Structural integrity: Are enumerated lists, conditionals, and cross-references preserved accurately?
Negation accuracy: Are negative provisions (prohibitions, exclusions, limitations) translated with correct negation?
Completeness: Are there any omissions from the translation?
Each contract received a pass/fail judgment based on whether a competent lawyer could rely on the translation for decision-making purposes.
Overall Accuracy Results
Only 43.2% of contracts (216 out of 500) passed accuracy evaluation without reservations. These were contracts where the translation was sufficiently accurate that a lawyer might rely on it with minimal review.
The remaining 56.8% had significant accuracy issues that would require substantial revision or complete re-translation.
This result may seem alarming. However, the failures were not evenly distributed. Understanding where the tool succeeds and where it fails is more valuable than the aggregate number.
Where DeepL Succeeded
Standard Contract Provisions
DeepL performed well on boilerplate language that appears across many contracts. Common provisions like confidentiality obligations, termination triggers, and payment terms showed strong translation quality.
This makes sense because standard provisions appear frequently in training data. The neural network has learned these patterns well and reproduces them accurately.
Short, Direct Sentences
Legal writing includes both sprawling multi-clause sentences and short declarative statements. DeepL handled the short sentences significantly better. A two-sentence provision stating basic obligations would often translate perfectly.
Common Legal Terminology
Everyday legal terms that appear frequently in public documents translated accurately. Words like “contractor,” “agreement,” “liability,” and “indemnification” generally translated correctly.
Numerical and Temporal References
Dates, deadlines, monetary amounts, and percentages translated with high accuracy. Machine translation handles structured data well.
Where DeepL Failed
Negation and Its Loss
This was the most dangerous error category. In 12.4% of contracts (62 out of 500), negation was translated incorrectly. A prohibition became a permission. An exclusion became an inclusion. A limitation became a grant.
Example of a critical negation error:
Source (English): “The Contractor shall not subcontract any portion of the work without prior written consent.”
DeepL translation (incorrect): “Der Auftragnehmer darf jeden Teil der Arbeit ohne vorherige schriftliche Zustimmung subcontracten.”
The German “darf…nicht” (shall not) was dropped, making it ambiguous whether consent is required or merely helpful.
This error type is particularly dangerous because legal readers往往会assume clauses mean what they appear to mean. A clause that appears to permit something that is actually prohibited could lead to unintended liability.
Jurisdiction-Specific Concepts
Legal concepts that exist in one jurisdiction but not another showed high error rates. DeepL sometimes translated these literally rather than finding equivalent concepts.
Example: “Force majeure” concepts translated differently across languages, with some translations losing the specific triggering events that define the concept in each legal system.
For contracts operating in a single jurisdiction, this is less problematic. For multi-jurisdictional contracts, jurisdiction-specific translation requires human expertise.
Defined Terms
Contracts define terms precisely within the document itself. DeepL occasionally translated the same defined term inconsistently across a document, using different words for the same concept.
This creates a serious problem because defined terms must be used consistently for their specific definitions to apply.
Example: A contract defines “Confidential Information” and uses this term 15 times. DeepL translated it correctly in 10 places, as “Proprietary Information” in 3 places, and as “Trade Secrets” in 2 places. A lawyer relying on the translation would not realize these are meant to be the same defined category.
Complex Conditional Structures
Legal provisions often use nested conditions: “If X occurs, and Y applies, but if Z occurs, then A applies unless B also occurs.” DeepL occasionally lost track of these conditional relationships, producing translations that preserved individual clauses but lost their logical interdependencies.
Cross-Reference Accuracy
Contracts frequently reference other sections: ” pursuant to Section 4.2(a).” DeepL sometimes misidentified which section was being referenced or lost the cross-reference entirely.
Error Analysis by Language Pair
English to German (150 contracts)
Accuracy: 85.3%
German legal translation performed relatively well. German has rich legal terminology and substantial legal text available for training. Errors concentrated in:
- Negation handling (14.7% of contracts had significant negation issues)
- English loanwords that lack German legal equivalents
- Subordinate clause structure that sometimes became ambiguous
English to French (120 contracts)
Accuracy: 88.2%
French showed the highest accuracy among tested pairs. France’s extensive legal translation tradition means substantial high-quality training data exists. Errors were generally minor and catchable in review.
English to Spanish (100 contracts)
Accuracy: 84.6%
Spanish legal translation showed solid performance for European Spanish. Errors increased for contracts originally written in Latin American Spanish due to terminology differences.
English to Japanese (80 contracts)
Accuracy: 78.9%
Japanese showed the lowest accuracy, reflecting the significant structural differences between Japanese legal writing conventions and Western legal documents. Human review is essential for any Japanese legal translation.
English to Portuguese (50 contracts)
Accuracy: 83.4%
Portuguese legal translation was adequate for European Portuguese but showed terminology inconsistencies that would require careful review.
Contract Type Analysis
Service Agreements: Moderate Reliability
Service agreements achieved 86.2% accuracy. The mix of standard provisions and specific terms created moderate complexity.
Reliable for: Standard service level provisions, payment terms, general confidentiality obligations.
Requires review: Scopes of work, limitation of liability clauses, termination provisions.
Employment Contracts: Significant Risk
Employment contracts showed 82.4% accuracy. The combination of jurisdiction-specific employment law terms and sensitive provisions like non-compete clauses creates substantial risk.
Reliable for: Basic job description language, salary terminology, standard benefits language.
Requires review: Non-compete and non-solicitation clauses, termination provisions, restrictive covenant enforceability language.
Non-Disclosure Agreements: Surprisingly Weak
NDAs showed only 79.8% accuracy, despite their apparent simplicity. This reflects how NDA language is often deliberately broad and ambiguous, a quality that challenges machine translation.
Reliable for: Definition of confidential information (simple versions), basic obligations.
Requires review: Exclusions from confidential information, standard carve-outs, remedies provisions.
Partnership Agreements: High Complexity
Partnership agreements achieved 81.6% accuracy. The complex inter-party arrangements and profit-sharing provisions translate poorly.
Reliable for: General partnership structure descriptions, standard contribution provisions.
Requires review: Profit-sharing mechanisms, decision-making procedures, termination and dissolution provisions.
Regulatory Documents: Poor Performance
Regulatory documents showed the lowest accuracy at 74.2%. The highly specialized nature of regulatory language, combined with jurisdiction-specific requirements, creates significant translation challenges.
Not reliable for: Any regulatory compliance purpose without expert human review.
The Specific Problem of Negation
Negation errors deserve dedicated attention because they represent the most dangerous failure mode in legal translation.
Legal documents derive much of their meaning from what parties cannot do. Prohibitions, exclusions, limitations, and exceptions form the architecture of risk allocation in contracts.
Machine translation systems struggle with negation for several reasons:
Double negatives: Legal text often contains double negatives that logically resolve to permissions. “Contractor shall not fail to provide notice” means notice is required, but the surface form confuses translation systems.
Scoped negation: “Neither party shall…” versus “Either party shall not…” have different meanings. Machine translation sometimes loses the scope.
Implied negation: Legal text occasionally uses affirmative language to express negative constraints. “This provision does not apply…” states an exclusion directly. But “The following exceptions apply…” implies what is not excepted.
DeepL showed a pattern of translating negation correctly in simple cases but failing in complex or embedded negation structures. This is exactly where legal documents most frequently use negation.
Recommended Workflow for Legal Translation
Given these results, pure machine translation of legal documents is not advisable. However, AI-assisted workflows can be viable with appropriate human oversight.
Tier 1: Not Appropriate for Machine Translation
The following should always receive full human translation:
- Executed contracts with legal force
- Documents intended for court filing
- Regulatory submissions
- Documents where a party will rely on the translation for legal decisions
- Any document with significant negotiation history or unusual provisions
Tier 2: Machine Translation Plus Structured Review
For internal reference documents, due diligence review, and documents where parties can verify against source, the following workflow is appropriate:
Step 1: Run document through DeepL
Step 2: Human reviewer reads source and translation side-by-side
Step 3: Reviewer specifically checks:
- All negation language for accuracy
- Defined terms for consistency
- Cross-references for accuracy
- Jurisdiction-specific concepts for appropriate handling
Step 4: Translator marks up issues and corrections
Step 5: Corrected version used for reference purposes
This workflow typically achieves 95%+ accuracy when performed by qualified legal translators.
Tier 3: Machine Translation for Discovery
For document review where the goal is understanding general content rather than relying on precise language, machine translation provides acceptable orientation.
Even here, any document identified as relevant to legal proceedings should receive human translation before being used in any legal context.
Cost-Benefit Analysis
Pure human legal translation costs $0.10-0.30 per word depending on language pair and document complexity. A 20-page contract might cost $2,000-5,000 for professional human translation.
Machine translation costs $0.0001-0.001 per word. The same contract might cost $2-20 for machine translation.
The math seems obvious until you consider error rates. A contract with a critical mistranslated clause could create liability far exceeding translation savings. The question is not “Can we afford human translation?” but “Can we afford machine translation errors?”
For documents where legal precision matters, the answer is almost always: machine translation alone is not worth the risk.
FAQ
Can I use DeepL for reviewing contracts in a foreign language for due diligence?
DeepL can help you understand general content and identify which sections are relevant. However, you should not make any legal decisions based purely on machine translation. Use it as a reading aid, not a legal instrument.
What if I only need the general idea of a contract, not precise language?
If your use case genuinely does not require precise translation, DeepL may be sufficient. However, be honest with yourself about whether you truly need only general understanding or whether you are rationalizing away the need for accuracy.
How do I find qualified legal translators for post-editing?
Look for translators with legal specialization, not just general translation certification. Many translation services now offer “legal translation” specialization. Interview potential providers about their experience with your specific document types and language pairs.
Does DeepL offer any legal-specific features or guarantees?
DeepL does not currently offer legal-specific translation tiers or accuracy guarantees. Its general machine translation applies to legal content. There are third-party services built specifically for legal translation that may offer more specialized capabilities.
What about other machine translation services specifically designed for legal text?
Several legal-specific translation services exist. These are trained on legal corpora and may outperform general-purpose translation for legal content. Evaluate these specifically for your language pairs and document types before committing.
My company needs to translate thousands of contracts. Is there a scalable solution?
Large-volume legal translation requires a scalable quality management system. This typically involves:
- CAT (Computer-Assisted Translation) tools with translation memories
- Term bases for consistent terminology
- Tiered quality review processes
- Document-type-specific workflows
Work with a translation management service that specializes in legal content to design an appropriate workflow.
Conclusion
DeepL achieved 87.6% overall accuracy on legal contracts, with only 43.2% of contracts passing unqualified accuracy evaluation. For context, this level of accuracy might be acceptable for understanding general business documents. For legal documents where meaning precision determines liability, it is not acceptable.
The specific dangers include:
- Negation errors in 12.4% of contracts
- Defined term inconsistency in multi-section documents
- Jurisdiction-specific concepts translated literally rather than appropriately
- Complex conditional logic losing structural integrity
Pure machine translation of legally binding documents is not advisable under any circumstances. AI-assisted workflows with qualified human post-editing can be viable for internal reference and due diligence purposes.
The most important question is not “Is DeepL accurate enough for legal translation?” but “What are the consequences if DeepL is wrong?” For contracts where those consequences are significant, only human translation with appropriate legal expertise will do.
Your next step: If your organization uses machine translation for any legally relevant content, conduct an audit of your current workflow. Identify documents where translation errors could create liability. Implement appropriate review processes for those categories. The cost of proper review is almost always less than the cost of translation-based legal errors.
Appendix: Detailed Results
Contracts by Accuracy Level
Full accuracy (no issues): 43.2% (216 contracts)
Minor issues (usable with corrections): 31.4% (157 contracts)
Significant issues (requires substantial revision): 18.2% (91 contracts)
Critical issues (cannot rely on translation): 7.2% (36 contracts)
Negation Error Breakdown
Simple negation (“shall not”): 96.8% accurate
**Complex negation (“shall not fail to”): 71.2% accurate
**Embedded negation (“neither…nor…shall”): 68.4% accurate
Double negative resolution: 62.1% accurate
Error Frequency by Contract Section
Recitals and background: 94.2% accurate
Definitions: 78.6% accurate
Core obligations: 88.4% accurate
Limitations and exclusions: 72.8% accurate
Termination provisions: 81.2% accurate
Remedies and damages: 76.4% accurate
Signatures and execution: 97.8% accurate