পাঠ ২৬ · ২৮-এর মধ্যে · মডিউল ৪
Home / AI Courses / Generative AI / Safety & Ethics

Safety, Bias ও Copyright সমস্যা

Safety & ethics — bias, deepfake, copyright, watermarking
৭ মিনিট পড়া সবার জন্য · All Responsible AI

এই পাঠে যা শিখবেন

  • Bias কীভাবে data → output amplify হয় ও measure
  • Deepfake-এর threat surface ও Bangladesh-এ documented case
  • Copyright debate — fair use vs infringement, ongoing lawsuits
  • Watermarking, C2PA, SynthID — provenance solution-এর strength ও limit

১ · Bias amplification — invisible training data

একটি ভয়াবহ truth: AI bias-free নয়। Training data-তে যা আছে — model তা amplify করে। Bias amplificationBias amplificationTraining data-র statistical bias output-এ stronger হয়ে আসে। Selection bias, loss function, scale সব contribute করে। documented across image, text, audio।

Examples:

  • Stable Diffusion "CEO" → ৯৫% white male।
  • "Beautiful person" → light-skinned female।
  • "Criminal" → dark-skinned male।
  • "Engineer" → male; "nurse" → female।
  • "Bangladeshi" prompt → poverty stereotype repeatedly।

২ · কোথা থেকে আসে

  • Web data bias: LAION, Common Crawl — Western, English-dominant।
  • Caption bias: stereotyping captions।
  • Aesthetic filtering: "high-quality" filter Eurocentric beauty।
  • Loss function: majority class easy → model lazy default।
  • RLHF labelers: demographic homogeneity।

৩ · Deepfake — threat & documented harm

২০২৪-২৫ election year — Bangladesh, India, US, UK — সব affected।

  • Non-consensual intimate imagery (NCII): ৯০%+ deepfake-এর target নারী। বাংলাদেশে cyber unit বহু case রিপোর্ট।
  • Voice clone fraud: "মা, আমার accident — টাকা পাঠাও" — global scam ছড়াচ্ছে।
  • Election misinformation: Tarique Rahman (BD), Imran Khan (PK), Biden robocall (US) — সব AI-generated।
  • Stock market manipulation: Pentagon explosion AI image — Wall Street ১% drop।
  • Job interview fraud: deepfake video resume, FBI 2022 warning।
Bangladesh Cyber Security Act 2023-এ deepfake-specific provision অস্পষ্ট। Defamation/obscenity-এর existing clause apply হলেও — burden of proof victim-এর উপর।

৪ · Copyright — historical lawsuits

  • Getty Images v Stability AI (2023): Stable Diffusion training-এ Getty image scraped — output-এ Getty watermark visible। UK ও US case pending।
  • Andersen et al. v Stability/Midjourney/DeviantArt (2023): artist class action — style mimicry।
  • NYT v OpenAI/Microsoft (Dec 2023): verbatim memorization evidence; billions claim।
  • Authors Guild v OpenAI (2023): Pulitzer winners include — books-এ training।
  • RIAA v Suno/Udio (June 2024): $150K per work damages।
  • Settlement (NYT, AP partial): some publishers license deal।

৫ · Studio Ghibli, artist style — gray zone

"Ghibli style" prompt → instantly recognizable। Style copyrightable নয় (US case law) — কিন্তু artist livelihood threatened।

  • ২০২৫ March — OpenAI GPT-4o image gen "Ghibli" style trend। Hayao Miyazaki past quote — "I am utterly disgusted"।
  • "Greg Rutkowski" — fantasy artist; SD-এ name removed (legal request)।
  • NIGHTSHADE (Glaze project) — artist tool যা training-এ image poison করে।

৬ · Watermarking — provenance solution

  • SynthID (Google DeepMind): imperceptible pattern in pixel/audio। Mid-2024 text-এও।
  • C2PA (Content Credentials): open standard — content metadata cryptographically sign। Adobe, Microsoft, Sony, Nikon, Leica adopting।
  • Metadata watermark: EXIF, XMP — easy strip।
  • Visual watermark: "AI-generated" overlay — easy crop।

৭ · Adversarial vulnerability

  • JPEG recompression, screenshot, crop, brightness — ৫০-৯০% watermark destroy।
  • "Adversarial purification" attack — explicitly remove।
  • Open-source model — watermark optional।
  • StegoFormer, Tree-Ring (NeurIPS 2023) — diffusion-native robust watermark research।
তিন প্রধান safety চ্যালেঞ্জ ⚠ Bias amplification "CEO" → 95% white male "nurse" → female-skewed Western beauty filter 🎭 Deepfake & misuse Election misinfo (BD 2024) Voice clone fraud NCII — 90% target women © Copyright disputes Getty v Stability NYT v OpenAI RIAA v Suno/Udio 🛡️ Multi-layer defense Data curation · RLHF safety · Output filter Watermark (SynthID) · C2PA provenance · Detector দায়িত্বশীলতা — সবাইকে 🏢 Model lab RLHF, red-team ⚖️ Legislator EU AI Act, US states 📱 Platform Meta, YouTube label 👤 User verify, disclose
তিন প্রধান চ্যালেঞ্জ — bias, deepfake, copyright। সমাধান layered + multi-stakeholder।

৮ · Bias measurement — code

Python · bias audit 
from diffusers import StableDiffusionPipeline
from deepface import DeepFace
import torch, collections

pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    torch_dtype=torch.float16,
).to("cuda")

# Generate 50 images per occupation
occupations = ["doctor", "nurse", "engineer", "teacher", "CEO"]
counts = {}
for occ in occupations:
    races, genders = [], []
    for i in range(50):
        img = pipe(f"a photo of a {occ}").images[0]
        img.save(f"/tmp/{occ}_{i}.jpg")
        try:
            r = DeepFace.analyze(f"/tmp/{occ}_{i}.jpg",
                                 actions=["gender", "race"],
                                 enforce_detection=False)[0]
            genders.append(r["dominant_gender"])
            races.append(r["dominant_race"])
        except: pass
    counts[occ] = (collections.Counter(genders),
                   collections.Counter(races))

for occ, (g, r) in counts.items():
    print(f"{occ}: gender={dict(g)}, race={dict(r)}")

    


    

      Reality: most SD models-এ এই audit-এ "engineer" 80%+ male, "CEO" 90%+ white male। DALL·E 3 অনেক আগে diversity prompt rewrite করে।
    


    
৯ · Bangladesh-specific concerns

    
    
      
  • Election: ২০২৪ election-এ AI voice/image circulated।
    • 
      
  • Cyber crime against women: deepfake porn complaint ৪x বৃদ্ধি (Police cyber unit data 2023-24)।
    • 
      
  • Misinformation channel: Facebook/WhatsApp viral Bangla deepfake — fact-check infrastructure weak।
    • 
      
  • Religious sensitivity: AI-generated religious imagery offense; communal tension trigger।
    • 
      
  • Bangladeshi face data: ID card, passport leak — voice/face training-এ misuse risk।
    • 
      
  • Cultural representation: "Bangladesh" prompt mainly poverty/disaster — orientalist bias।
    • 
    


    

      Bangladesh-এ Personal Data Protection Act (draft 2023) ও Digital Security Act revision — AI-aware provision চাই। Civil society advocacy critical।
    


  


  

    
    
  


  

    
 ভাবনার প্রশ্ন


    

      
        প্র ০১
        "Bias mitigation" — debiasing, prompt rewriting, fine-tune। কোনটা কাজ করে? Google Gemini-র "diverse Founding Fathers" বিতর্ক থেকে কী শিখি?
      
      

        
২০২৪ Feb-এ Google Gemini "George Washington" → black/Asian rendered। Backlash। ২ সপ্তাহ pause। Bias mitigation-এর complexity exposed।

        
Approaches:

        
    
          
  • Data debiasing: training set demographic balance। ভাল কিন্তু expensive ও subjective।
    • 
          
  • Prompt rewriting: "doctor" → "doctor, diverse race and gender" silently inject। Gemini-র failure mode।
    • 
          
  • RLHF safety tuning: "depict diverse subjects" reward। Over-correct risk।
    • 
          
  • Fine-tune on balanced subset: targeted update।
    • 
          
  • Counterfactual augmentation: training-এ flipped attribute add।
    • 
          
  • Conditional generation: user-controlled diversity slider।
    • 
        

        
Gemini case lessons:

        
    
          
  • Historical accuracy violation: "Roman emperor" black female — historically wrong।
    • 
          
  • Naive prompt injection: system-level rewrite — context-blind।
    • 
          
  • Cultural insensitivity: "Vikings" non-white — original distort।
    • 
          
  • Unintended consequence: over-correction = new bias।
    • 
        

        
Tradeoffs:

        
    
          
  • Generic queries (CEO, nurse): diversity injection valid।
    • 
          
  • Historical/specific: accuracy primary।
    • 
          
  • Cultural context: context-aware modeling।
    • 
        

        
Better approach:

        
    
          
  • Intent classifier — historical vs hypothetical।
    • 
          
  • User control — explicit demographic specification।
    • 
          
  • Disclosure — "I diversified the depiction"।
    • 
          
  • Quality gates — historical accuracy benchmark।
    • 
          
  • Iterative — not single binary fix।
    • 
        

        
Empirical results:

        
    
          
  • DALL·E 3 silent rewriting — controversial but sometimes helpful।
    • 
          
  • SDXL — base bias high; ControlNet/conditioning user-side।
    • 
          
  • FLUX (২০২৪) — improved but bias remain।
    • 
        

        
Bangladesh context:

        
    
          
  • "Bangladeshi family" → poverty-laden imagery default।
    • 
          
  • "Beautiful woman" → light skin bias।
    • 
          
  • "Wedding" → Indian-Western mixed, not Bangla specific।
    • 
          
  • Fine-tune on Bangladeshi imagery — community-led আবশ্যক।
    • 
        

        
মূল উপলব্ধি: Bias mitigation context-dependent। Single magic bullet নেই। Transparency + user control + iterative measurement essential। Gemini-র mistake — well-intentioned but technically/culturally naive।

      

    


    

      
        প্র ০২
        NCII (deepfake porn) — Bangladesh-এ এপিডেমিক। Tech, legal, support — কী stack চাই? Take-down speed কতটা important?
      
      

        
NCII — non-consensual intimate imagery — woman ও girl সবচেয়ে affected। Cyber unit-এ Bangladesh complaint বছরে ৪x বাড়ছে।

        
Tech defense:

        
    
          
  • 
            Hash matching (StopNCII.org, NCMEC): victim image hash → platform matches → take-down। Photo never upload — only hash।
          
    • 
          
  • 
            Detection model: deepfake classifier ৭০-৯০% accuracy। Adversarially fragile।
          
    • 
          
  • 
            Generation prevention: face-aware filter at generation; identity-blocking ("don't generate real people")।
          
    • 
          
  • 
            Search engine demotion: Google, Bing — query/result filter।
          
    • 
          
  • 
            Watermark forensics: detect AI-origin।
          
    • 
        

        
Legal landscape:

        
    
          
  • 
            UK Online Safety Act 2023: NCII illegal regardless of consent at creation।
          
    • 
          
  • 
            UK 2024 — deepfake porn creation criminal: not just sharing।
          
    • 
          
  • 
            US — TAKE IT DOWN Act (২০২৫): federal criminalization, ৪৮-hour platform take-down requirement।
          
    • 
          
  • 
            South Korea: harsh sentences after 2024 epidemic।
          
    • 
          
  • 
            Bangladesh: Cyber Security Act 2023 obscenity/defamation cover, কিন্তু deepfake-specific provision absent। Pornography Control Act 2012 outdated।
          
    • 
        

        
Take-down speed importance:

        
    
          
  • প্রথম ২৪ ঘণ্টায় 80% damage — viral spread।
    • 
          
  • Once on Telegram/4chan — irretrievable।
    • 
          
  • SLA target: < ৪৮ hours (UK gold standard); < ২৪ hours (best)।
    • 
          
  • Platform varied — Meta-র mixed record; Telegram/X-এ slower।
    • 
        

        
Survivor support:

        
    
          
  • Cyber Police Bangladesh, BdCSIRT: reporting channel।
    • 
          
  • Mahila Parishad, Naripokkho: civil society।
    • 
          
  • Mental health: Kaan Pete Roi, Moner Bondhu — Bangla support।
    • 
          
  • Legal aid: Bangladesh Legal Aid (BLAST), Ain o Salish Kendra।
    • 
          
  • Reputation cleanup: SEO suppression, hash-and-take-down service।
    • 
          
  • Family/social shame: often biggest barrier — community education।
    • 
        

        
Education / prevention:

        
    
          
  • School curriculum — digital safety।
    • 
          
  • Parental awareness।
    • 
          
  • Photo-sharing hygiene।
    • 
          
  • Bystander intervention training।
    • 
        

        
Platform accountability:

        
    
          
  • Bangladesh BTRC — pressure on Facebook, YouTube, TikTok।
    • 
          
  • Local language moderation — Bangla content moderator chronically understaffed।
    • 
          
  • Transparency report mandatory।
    • 
        

        
Multi-stakeholder coalition needed:

        
    
          
  • Government — law update।
    • 
          
  • Tech companies — detection + take-down।
    • 
          
  • NGO — survivor support।
    • 
          
  • Media — responsible reporting।
    • 
          
  • Schools — prevention।
    • 
        

        
মূল উপলব্ধি: NCII-এ pure tech solution নেই — sociotechnical। Bangladesh-এ urgent legislative + capacity building। Speed, support, prevention — তিনটিই simultaneously দরকার।

      

    


    

      
        প্র ০৩
        "Fair use" vs copyright — AI training-এ কে জিতবে? Stability v Getty, NYT v OpenAI court case-এর তাৎপর্য কী?
      
      

        
২০২৫-এ AI copyright-এর precedent তৈরি হচ্ছে। Outcome generation industry-র shape define করবে।

        
Fair use 4 factor (US):

        
    
          
  • (১) Purpose & character — transformative? commercial?
    • 
          
  • (২) Nature of work — factual / creative?
    • 
          
  • (৩) Amount used — entire work? portion?
    • 
          
  • (৪) Market effect — substitution risk?
    • 
        

        
OpenAI/Stability defense:

        
    
          
  • Training "transformative" — output ≠ input।
    • 
          
  • Authors Guild v Google Books (2014) precedent — full-text scan ruled fair use।
    • 
          
  • "Statistical extraction" — like reading ও learning।
    • 
          
  • Public benefit (research, education)।
    • 
        

        
Plaintiff arguments:

        
    
          
  • Verbatim memorization evidence (NYT exhibit) — direct reproduction।
    • 
          
  • Market substitution — AI summary erodes news subscription।
    • 
          
  • Style mimicry — artist livelihood threat।
    • 
          
  • Watermark/credit removed (DMCA 1201 violation)।
    • 
        

        
Recent rulings (2024-25):

        
    
          
  • Thomson Reuters v Ross Intelligence (Feb 2025): first AI training fair use ruling — Reuters won। But specific to legal headnote, not generative AI।
    • 
          
  • Andersen v Stability: Surviving motion to dismiss — case proceeding।
    • 
          
  • NYT v OpenAI: discovery phase ongoing।
    • 
          
  • OpenAI/Anthropic settlement attempts: some publisher license।
    • 
        

        
Possible outcomes:

        
    
          
  • Outcome A — fair use upheld: AI industry continues; voluntary license norm।
    • 
          
  • Outcome B — copyright infringement found: training data licensing requirement; cost spike; smaller players excluded।
    • 
          
  • Outcome C — partial — case-by-case: verbatim bad, statistical OK।
    • 
          
  • Outcome D — legislative intervention: Congress/Parliament-এ statutory license।
    • 
        

        
International divergence:

        
    
          
  • EU AI Act — training data transparency mandatory; opt-out provision।
    • 
          
  • UK — text-and-data-mining exception under review।
    • 
          
  • Japan — AI training fair-use friendly (2018 amendment)।
    • 
          
  • China — "national interest" framing, looser।
    • 
          
  • Bangladesh — Copyright Act 2000 + 2023 amendment — AI provision absent। DPDC court next decade decision-এ depend।
    • 
        

        
Industry response:

        
    
          
  • Licensed datasets boom (Shutterstock-OpenAI, Reddit-Google deal)।
    • 
          
  • Synthetic data generation।
    • 
          
  • "Opt-out" registry (Spawning, HaveIBeenTrained)।
    • 
          
  • Data provenance standards।
    • 
        

        
Bangladesh-specific:

        
    
          
  • Bangla newspaper, literature — protection mechanism।
    • 
          
  • Tagore — public domain (Indian copyright expired); but performance/translation rights।
    • 
          
  • Local artists — class action infrastructure absent।
    • 
        

        
মূল উপলব্ধি: Outcome uncertain কিন্তু industry maturity-এ এই lawsuits inevitable। "Fast & loose" era ending; license + provenance era beginning। Bangladesh-এর policy-makers এখন থেকেই think করা উচিত।

      

    


    

      
        প্র ০৪
        Watermarking (SynthID, C2PA) — adoption challenge কী? "Bad actor"-রা use করবে না — তবু worth-it কেন?
      
      

        
Watermarking ও provenance — silver bullet নয়, কিন্তু important layer।

        
SynthID (Google) approach:

        
    
          
  • Image — pixel-level imperceptible pattern।
    • 
          
  • Audio — frequency domain signature।
    • 
          
  • Text (mid-2024) — token sampling-এ subtle bias।
    • 
          
  • Detection — Google-only access।
    • 
          
  • Robust to JPEG, screenshot, slight crop।
    • 
          
  • Brittle to heavy edit, regeneration।
    • 
        

        
C2PA (Content Credentials) approach:

        
    
          
  • Open standard — Adobe, Microsoft, OpenAI, Sony, Nikon, Leica adopting।
    • 
          
  • Cryptographic signing — tamper detection।
    • 
          
  • Edit history — provenance chain।
    • 
          
  • Browser extension — Content Credentials display।
    • 
          
  • Strip-easy — but mismatch detectable।
    • 
        

        
Bad actor problem:

        
    
          
  • Open-source model — watermark removable।
    • 
          
  • Re-encode পরে — destroyed।
    • 
          
  • Adversarial attack — explicit removal।
    • 
          
  • Russia, China state actor — won't comply।
    • 
          
  • Telegram, niche chans — won't moderate।
    • 
        

        
Why still worth-it:

        
    
          
  • Default authenticity: good actor (newspaper, brand) sign content। Unsigned content = suspicion default।
    • 
          
  • Lawsuit evidence: "this image carries Adobe + OpenAI signature → AI-origin proven"।
    • 
          
  • Platform policy: Meta, YouTube — AI label or take-down।
    • 
          
  • Ecosystem creation: camera (Nikon, Sony) embed at capture; reporter use।
    • 
          
  • Election integrity: verified candidate channel; unverified = doubt।
    • 
          
  • Court admissibility: signed evidence chain।
    • 
        

        
Adoption challenges:

        
    
          
  • Network effect: few adopt → low value।
    • 
          
  • UI clutter: per-image badge fatigue।
    • 
          
  • Privacy concern: camera signature → device tracking?
    • 
          
  • Legacy content: pre-2024 unsigned forever।
    • 
          
  • Open-source model: Stable Diffusion no built-in watermark।
    • 
          
  • Cost: tooling, licensing, certificate authority।
    • 
        

        
Complementary defense:

        
    
          
  • Detection model (FakeCatcher, Hive, Optic) — adversarial arms race।
    • 
          
  • Source-side verification — politician official channel।
    • 
          
  • Reverse-image search — earliest occurrence track।
    • 
          
  • Crowd-sourced fact-check (Community Notes)।
    • 
          
  • Media literacy education।
    • 
        

        
Bangladesh adoption strategy:

        
    
          
  • BTRC — platform Bangla content moderation pressure।
    • 
          
  • Election Commission — candidate official channel + signature।
    • 
          
  • Major media (Prothom Alo, BBC Bangla) — C2PA pioneer।
    • 
          
  • BdCSIRT — citizen detection toolkit Bangla-language।
    • 
          
  • School digital literacy curriculum।
    • 
        

        
Future direction:

        
    
          
  • Tree-Ring, Stable Signature — diffusion-native watermark research।
    • 
          
  • Hardware attestation — phone camera signed boot।
    • 
          
  • Decentralized provenance (IPFS + signature)।
    • 
        

        
মূল উপলব্ধি: Watermark perfection-এর জন্য নয় — friction-এর জন্য। ৮০% bad actor sloppy — friction them। ২০% sophisticated — other defense layer। Imperfect tool useful while perfect tool elusive।

      

    

  


  

    
 অনুশীলন

    
    
      
  1. 
        Bias audit: "professional doctor" ও "professional nurse" prompt-এ ১০টি SD image generate করে gender ratio note করুন।
        
    
          
          
    
            
    Typical SD-1.5: doctor → ৮০-৯০% male, nurse → ৮০-৯০% female। Bias clear। SDXL slight improvement; FLUX best-of-OS।
    
          
    
        
    
      
    2. 
      
  2. 
        Detection: একটি AI image (SD generate) ও একটি real photo নিয়ে — বিনামূল্যের detector (sightengine, hivemoderation, isitai.com) test করুন।
        
    
          
          
    
            
    Detector accuracy varied — 70-90% in controlled, 40-70% in-the-wild। Confidence interpretation careful।
    
          
    
        
    
      
    3. 
      
  3. 
        ভাবুন: Bangladesh election commission-এর জন্য একটি deepfake response protocol design করুন — ৫টি স্টেপ।
        
    
          
          
    
            
      
              
    • (১) Verified channel — candidate official C2PA-signed video।
      • 
              
    • (২) ২৪x৭ rapid response unit — citizen tip → analyst।
      • 
              
    • (৩) Platform partnership — Meta, YouTube, TikTok-এ priority take-down।
      • 
              
    • (৪) Public detection portal — citizens upload, get verdict।
      • 
              
    • (৫) Media literacy campaign — Bangla TV/radio public service।
      • 
            
    
          
    
        
    
      
    4. 
    

  


  

    
 আরও পড়ুন

    
    

      Resource: StopNCII.org — free hash-based image take-down। ভিকটিম support: Content Credentials।
    

  







  
   এই পাঠে ভুল পেয়েছেন? জানান





  
    
 পূর্ববর্তী পাঠ

    
পাঠ ২৫ · Evaluation