As large language models (LLMs) become vital to business and in your personal life, we must know which LLMs align with Ethical AI principles.
​
This is key for their responsible use.
​
This analysis tests major AI models against Human-Centred Trustworthy AI Principles. It provides evidence-based insights, not marketing claims.

​
Why This Assessment Matters

Your choice of LLM isn’t just about its tech. It’s about trust, responsibility, and ethics. These principles help you make choices that protect you as a end-user and your organisation.
​
So let’s dive in.

Meta’s Llama
Total Score: 18/80

Critical failures across multiple principles reveal significant concerns. The model has faced massive data breaches of 530 million Facebook accounts.
​
It also violated GDPR rules, leading to record fines for privacy breaches.
​
Key Issues:
- Cannot distinguish between sensitive and regular data
- Multiple privacy violation lawsuits
- Discriminatory ad-targeting practices
- Unfair “pay-or-consent” model
- Poor data protection practices
​

​

Riley’s take:
Wouldn’t touch it with a barge pole mate.
​
​Also recommend anyone outside EU check with a Facebook Account to check the opt-out options in your location for Meta using your data for training.
​
​

Grok
​Total Score: 19/80

​
XAI’s Grok demonstrates concerning gaps across multiple principles. Its “anti-woke” stance and use of Twitter data raise concerns about bias.
​
Critical Issues:
- Training on X/Twitter data inherently incorporates platform biases
- Lack of bias detection or mitigation systems
- Poor data protection practices
- No independent oversight
- Limited monitoring systems
​
​

​

Riley’s take:
Only consider using if you prefer your data privacy breached and enjoy having biassed AI.

Microsoft Copilot
Total Score: 40/80

​

Has mixed performance. It has good bias detection tools. But, it has major issues with basic functions and privacy.
​

Notable Aspects:
- Performance issues with larger datasets
- Security concerns regarding overly permissive data access
- Limited explanation of system decisions
- Implementation of InterpretML and Fairlearn for bias detection
- Significant usability issues reported

​

​

​

Riley’s take:
Genuinely surprised by this one. Being an Enterprise tool I would have expected a more robust framework and adherence to Ethical AI.

​

ChatGPT
Total Score: 50/80

​

It shows a steady, moderate level of performance across principles. It is strong in privacy protection but could improve in transparency.
​
Key Features:
- Clear data governance policies
- Regular model updates based on feedback
- Implementation of content filtering systems
- Challenges with attribution and sources
- Limited explanation of decision-making processes
​
​

​

Riley’s take:
Given its the most popular tool, its surprisingly mediocre when it comes to being Ethically built.

​
​

Gemini
Total Score: 59/80

​

Strong performance across principles, especially in safety and compliance, with robust governance.
​
Key Strengths:
- Strong data protection measures
- Rigorous review processes and ethical principles
- Clear commitment to regulatory compliance
- Dedicated teams for fairness and bias mitigation
- Comprehensive safety measures
​
​
​

Riley’s take:
Good to see it far outperforms Microsoft on this in enterprise space, but more needs to be done.

​

Free Webinar

Wednesday 6th November ​
​

Build Your AI Literacy Flywheel Are you ready to become a confident AI-enhanced professional with an unfair advantage of your peers? ​ Why AI Literacy is essential for everyone The 7 components of AI literacy My playbook becoming AI literate And so much more!

Date Wednesday 6th November TIME 9:30 CET ​(Europe Central) or 19:30 AEDT ​ (Aust East Coast) Location Online (Zoom) Level Beginners onwards Register Now

Mistral AI
Total Score: 62/80

​

Demonstrates impressive performance particularly in robustness and transparency through its open-source approach.

​

Notable Achievements:
- Open-source development fostering transparency
- EU-hosted infrastructure ensuring GDPR compliance
- Strong performance on standard benchmarks
- Clear licensing framework
- Multiple deployment options for security

​

​

​

​

Riley’s take:
As the only entrant from EU, it’s clearly takes responsible ethical AI seriously.

Claude
Total Score: 64/80

​

As the current frontrunner in Human-Centred Trustworthy AI implementation, Claude stands out through its Constitutional AI approach - embedding ethical guidelines directly into its training rather than adding them as afterthoughts.

​

Key Strengths:

🔒 Privacy & Security (9/10)

• 30-day auto-deletion of conversations
• No training on user interactions
• Strong data minimization practices
• Clear data handling policies
  ​

🛡️ Safety & Moderation (9/10)

• Sophisticated content filtering
• Nuanced understanding of harmful content
• Real-time output moderation
• Strong stance against malicious use
  ​

👤 Human-Centric Design (8/10)

• Adaptive communication style
• Clear acknowledgment of limitations
• Step-by-step explanations
• Focus on augmenting human capabilities
  ​

Built-in Guardrails:

• Clear knowledge boundaries (April 2024 cutoff)
• Strong ethical boundaries (won't assist with harmful content)
• Professional behavioral framework
• Robust data protection practices

​

​

​

Riley’s take:

Not at all surprised it’s the top - it’s the reason it’s my primary choice - having ethics has been built into the core.
​
Highly recommend - their data privacy and ethical stance is something you should really consider. It is also one of the best performing LLMs as well.

​
​

Making Informed Decisions

​

When selecting an LLM for yourself or your organisation, consider:
​
1. Risk Profile Matching

• High-risk sectors (healthcare, finance): Consider Claude or Gemini
• Open-source requirements: Look at Mistral AI
• General business use: ChatGPT or Copilot may suffice

​
2. Priority Alignment

• Privacy & Compliance: Claude, Mistral AI, or Gemini
• Transparency & Openness: Mistral AI
• Balanced Performance: Claude or Gemini

​

​

With all this in mind you might want to rethink which LLM you decide to use.
​
If so, before adopting any LLM, ask yourself:

1. What are your non-negotiable ethical requirements?
2. How will you monitor and assess AI outputs?
3. What additional safeguards might you need?

​

​

Remember: Even the highest-scoring models require human oversight and regular assessment. The goal isn’t perfection. It’s to implement things responsibly.
​
They must align with your organisation’s values and ethics.

​

​

I hope this has been helpful.

​

Riley
​
​