Skip to content

Responsible AI & Risk Management

Critical Principle

Access to AI tools does not require data exposure.

You can leverage AI's power whilst protecting AmaDema's intellectual property.

Understanding Hallucinations

What Are Hallucinations?

Hallucination: When an LLM generates plausible-sounding but factually incorrect or fabricated information.

Why Do They Occur?

LLMs are pattern-matching prediction engines, not knowledge databases.

When you ask:

"What is the tensile strength of Ti₆Al₄V processed at 400°C with 2% oxygen content?"

The model doesn't:

  • ❌ Look up this value in a database
  • ❌ Calculate it from first principles
  • ❌ Admit uncertainty

Instead, it:

  • ✅ Predicts the most probable next tokens based on training patterns
  • ✅ Synthesises a plausible-sounding answer from similar contexts
  • ✅ Presents it with confidence

Types of Hallucinations

1. Confident Fabrication

Example:

Prompt: "What is the melting point of polyethylene glycol-alt-polylactic acid copolymer with 40% PEG content?"

Response: "The melting point is 87.3°C, as reported by Zhang et al. (2019) in Polymer Science Journal 45(3):234-241."

Problem:

  • Specific temperature seems precise
  • Citation appears legitimate
  • But: Paper doesn't exist, value is fabricated

2. Plausible Interpolation

Example:

Prompt: "What is the Young's modulus of PLA nanofibers with 7% graphene?"

Response: "Approximately 4.2 GPa, based on typical reinforcement trends for graphene-loaded PLA."

Problem:

  • Sounds reasonable (graphene does increase modulus)
  • "Approximately" creates false credibility
  • But: Actual value depends heavily on dispersion, fiber alignment, testing conditions
  • Could be anywhere from 2-8 GPa

3. Non-Existent Citations

Example:

Prompt: "Find papers on non-oxide ceramic nanofibers."

Response: Lists 10 papers with authors, titles, journals, DOIs.

Problem:

  • 3-4 are real papers
  • 6-7 are fabricated (plausible combinations of real authors, realistic titles, fake DOIs)
  • Verify every citation

4. Misattributed Knowledge

Example:

Prompt: "Who discovered carbon nanotubes?"

Response: "Richard Smalley discovered carbon nanotubes in 1991."

Problem:

  • Smalley worked on fullerenes (different structure)
  • Iijima is generally credited with CNT discovery (1991)
  • Model conflated related research

The Red List: Data You Must Never Share

The Red List Protocol

These data types must NEVER be uploaded to public AI models (ChatGPT, Claude, Gemini, etc.)

Category 1: Unpublished Research Data

🚫 Novel molecular structures (before patent filing)
🚫 Exact synthesis parameters for proprietary processes
🚫 Experimental results from ongoing R&D
🚫 Failed experiments (negative data has IP value)
🚫 Grant applications under review

Category 2: Commercial Sensitive Information

🚫 Exact formulations (precursor ratios, additives)
🚫 Process temperatures/pressures for proprietary methods
🚫 Yield data that reveals manufacturing efficiency
🚫 Cost breakdowns and supplier information
🚫 Customer identities and partnership details
🚫 Pricing strategies and profit margins

Category 3: Personal & Confidential Data

🚫 Employee information (names, contact details, salaries)
🚫 Customer data (contacts, orders, communications)
🚫 Internal communications (emails, meeting notes with strategy)
🚫 Financial data (budgets, forecasts, bank details)
🚫 Legal documents (contracts, NDAs, IP correspondence)

Category 4: Security-Sensitive Information

🚫 Access credentials (passwords, API keys, tokens)
🚫 System configurations (server setups, network architecture)
🚫 Security protocols (physical access, data backup procedures)
🚫 Vulnerability assessments and penetration test results

Data Sanitisation Strategies

You can still use AI for sensitive tasks by sanitising data before processing.

Strategy 1: Anonymisation

Replace specific details with generic placeholders.

❌ Original (DO NOT SHARE): 

We synthesised PLA/graphene nanofibers using electrospinning at 
25kV with a 15cm working distance. Precursor solution: 12% PLA 
(Mw 100kDa) in DMF:DCM (3:1), with 5% graphene oxide reduced 
in-situ using hydrazine vapour at 80°C for 4 hours. Yield: 87%.

✅ Sanitised (SAFE TO SHARE): 

We synthesised polymer/nanofiller composite nanofibers using 
electrospinning at [VOLTAGE] with [DISTANCE] working distance. 
Precursor solution: [CONCENTRATION]% polymer (Mw [VALUE]) in 
[SOLVENT], with [X]% nanofiller processed via [METHOD]. 
Yield: [HIGH/MEDIUM/LOW].

Now you can ask:

"Convert this generic protocol description into a standardised template with sections for Materials, Methods, and Characterisation."

Strategy 2: Aggregation

Share trends rather than specific data points.

❌ Specific (Red List violation): 

Synthesis batch #343: Tensile strength 45.2 MPa at 23°C, 
42.1 MPa at 40°C, 38.9 MPa at 60°C

✅ Aggregated (Safe): 

Tensile strength decreases approximately 15% between 20-60°C 
for our polymer composite

Strategy 3: Hypothetical Framing

Ask about general principles, not your specific case.

❌ Specific (reveals IP): 

We're using in-situ reduction during electrospinning to disperse 
graphene oxide in PLA. What synthesis parameters should we optimise?

✅ Hypothetical (Safe): 

For in-situ reduction of nanofillers during polymer processing 
(generic question), what parameters typically affect dispersion 
quality in polymer nanocomposites? Provide a general framework 
for optimisation, not specific values.

Strategy 4: Use Local Models

For truly sensitive work, use the air-gapped sandbox.

The local Llama model:

  • ✅ Your data never leaves the room
  • ✅ No internet connection required
  • ✅ No storage by external providers
  • ✅ Full admin monitoring

Best for:

  • Analysing proprietary experimental data
  • Drafting patent applications
  • Processing customer information
  • Strategic planning documents

Verification Protocols

Never trust AI outputs without verification.

Protocol 1: Citation Verification

For every citation:

  1. ✅ Verify DOI resolves to real paper
  2. ✅ Check authors match
  3. ✅ Confirm journal and year
  4. ✅ Read abstract—does claim match?
  5. ✅ Check actual data in paper

Use: CrossRef, Google Scholar, PubMed

Protocol 2: Quantitative Data Verification

For any numerical claim:

  1. ✅ Check if value is within physically reasonable range
  2. ✅ Compare to known benchmarks for similar materials
  3. ✅ Verify units are consistent
  4. ✅ Look up value in authoritative source (handbook, database)
  5. ✅ If critical: independently calculate or measure

Red flags: - Suspiciously round numbers (exactly 100°C, 5.0 GPa) - Excessive precision (87.34256°C—unrealistic measurement precision) - Values outside known ranges (PLA melting at 300°C—impossible)

Protocol 3: Logical Consistency Check

Ask: - Does this conclusion follow from the premises? - Are there internal contradictions? - Does it contradict established knowledge? - Would an expert in this field agree?

Example:

AI Output: "Increasing graphene content from 5% to 10% improved tensile strength by 40% while simultaneously increasing elongation at break by 30%."

Red flag: Typically, higher filler content increases stiffness (tensile strength) but decreases ductility (elongation). This simultaneous improvement is possible but unusual—requires verification.

Protocol 4: The "Skeptical Colleague" Test

Before using AI output:

"If I presented this to my most skeptical colleague, what would they challenge?"

Then:

  1. Pre-emptively verify those points
  2. Add disclaimers where uncertainty exists
  3. Provide supporting evidence for critical claims

The Hallucination Hunt Exercise

Challenge: Find the Errors

You'll receive an AI-generated synthesis report for a PLA/graphene nanocomposite.

Hidden errors (5 total): - 1 fabricated citation - 1 physically impossible value - 1 misattributed discovery - 1 inconsistent claim - 1 non-existent characterisation method

Time limit: 20 minutes

Prize: "Critical Thinker" badge for fastest team

Sample Report Excerpt

"The nanocomposite was prepared following the method of Zhang et al. (2023, Polymer Engineering, DOI: 10.1016/j.poleng.2023.05.234). Graphene oxide was dispersed in N,N-dimethylformamide via sonication for 30 minutes, then mixed with 12% w/v PLA solution (Mw 150 kDa). The mixture was electrospun at 25 kV with a working distance of 15 cm. In-situ reduction was achieved using hydrazine vapour at 120°C for 2 hours, following the protocol first developed by Smalley (1996) for carbon nanotube functionalisation.

Characterisation via nano-X-ray photoelectron spectroscopy (nano-XPS) confirmed complete reduction of graphene oxide, with C/O ratio increasing from 2.1 to 15.8. Tensile testing showed Young's modulus of 4.2 GPa at 5% graphene loading, increasing to 8.9 GPa at 10% loading, while maintaining elongation at break >150% for both compositions. The melting point of the nanocomposite was measured at 168.5°C via DSC, consistent with the known value for PLA."

What's wrong? (Answers in workshop—no spoilers here!)

Red List Violations: Case Studies

Case Study 1: The Patent Disaster

Scenario: A researcher asked ChatGPT to help draft a patent application, including the complete synthesis protocol for a novel ceramic nanowire.

What happened:

  1. Protocol was processed by OpenAI's servers
  2. Data potentially used for model training (uncertain)
  3. Timestamp creates "prior disclosure" risk
  4. Patent application delayed 6 months for legal review
  5. Had to file provisional patent immediately

Lesson: Never share unpublished IP with public models.

Case Study 2: The Competitor Leak

Scenario: An R&D manager uploaded internal quarterly report (including customer names, pricing, and roadmap) to Claude for "executive summary generation."

What happened:

  1. Document contained customer NDAs
  2. Violated data protection agreements
  3. Potential GDPR violation (customer data)
  4. Required disclosure to affected customers
  5. Damaged business relationships

Lesson: Sanitise all internal documents before processing.

Case Study 3: The Successful Approach

Scenario: A materials scientist needed help analysing SEM images for fiber diameter distribution.

What they did:

  1. Used local sandbox (air-gapped)
  2. Processed images entirely offline
  3. Generated Python script for analysis
  4. Verified calculations independently
  5. No external data sharing

Outcome: ✅ Faster analysis, full IP protection, zero risk.

Lesson: For sensitive data, use local models.

Decision Framework: Can I Share This?

Is this data published or public?
├─ YES → Safe to use with any AI tool
└─ NO → Is it on the Red List?
    ├─ YES → Use local sandbox ONLY
    └─ NO → Can I sanitise it (anonymise, aggregate, hypothetical)?
        ├─ YES → Sanitise, then use external AI
        └─ NO → Use local sandbox ONLY

Exercise: Red List Assessment

Challenge

For each scenario, decide: ✅ Safe to share, ⚠️ Safe after sanitisation, or 🚫 Red List violation

  1. "Analyse this SEM image to measure fiber diameters" [image attached from ongoing experiment]
  2. "Summarise these 5 published papers on PLA degradation" [DOIs provided]
  3. "Convert this lab notebook entry into standardised format" [contains exact synthesis conditions for patent-pending process]
  4. "Generate Python code to calculate Young's modulus from stress-strain curves" [generic request, no data attached]
  5. "Draft an email to our customer Aerospace Corp about delay in shipment" [mentions company name]

Discuss: What sanitisation would make unsafe items safe?

Next: Day 1 Exercises: Put frameworks into practice →