AI Staged the Room. The Law Wants the Before Photo. Build the Oppy That Keeps Both.

AI Staged the Room. The Law Wants the Before Photo. Build the Oppy That Keeps Both.

By Anna with Oppy

A New York bill introduced on July 29 would require a conspicuous disclosure beside every materially altered or AI-generated real-estate image used to sell or lease property. Online, the broker, seller, owner, or property manager would also have to include the unaltered version or a link to it. The bill is in the Assembly Housing Committee. It is a proposal, not a law.1

California is not waiting for the hypothetical. Its AB 723 already requires a broker, salesperson, or person acting on their behalf to place a reasonably conspicuous disclosure on or next to a digitally altered sales image, then provide a public link, URL, or QR code to the original.2

The practical rule is becoming clear: an edited property image is no longer one file. It is a relationship between two files, one disclosure, one publishing destination, and one accountable human.

Most marketing workflows track only the pretty file.

That is the bug.

The disclosure is not the hard part

An agent can type “virtually staged” in five seconds. The difficult part arrives three weeks later when someone asks five less decorative questions.

Which original produced this image? What changed? Was the edit ordinary color correction or a new floor pretending to exist? Did the MLS require the original immediately before or after the altered version? Did the social post preserve the label after resizing? Did the property manager publish a lease advertisement under a rule written for sales?

This is not a prompt problem. It is a provenance and distribution problem.

On August 24, RE Technology summarized a new interview with Real Estate Standards Organization CEO Sam DeBord by asking whether data standards matter when AI can understand almost anything. Its answer was succinct: they “may matter more than ever.”3

They do. AI can inspect an image. It cannot recover an original that nobody retained, infer a destination rule that nobody encoded, or prove that the correct pair reached the public.

Boston Consulting Group described the same failure pattern in another industry on August 24. Automate one step while leaving the next one manual and “the bottleneck simply moves to the part of the workflow that AI hasn't touched.”4

Generate, disclose, publish, verify. Miss one verb and the workflow remains a demo.

One country, several operating rules

The direction is consistent. The mechanics are not.

The table is the reason a universal “AI edited” watermark is not a complete system. California requires access to the original. CRMLS specifies adjacency and warns first, then may impose a $250 fine if the issue is not corrected promptly. NorthstarMLS draws a line between personal property, such as furniture, and permanent features, such as walls, flooring, roofing, windows, and grading.5 6

NAR’s legal guidance offers the useful human version. Before publishing an AI-enhanced property image, ask whether the edit improves the client experience or attempts to “pull a fast one.”7

Software needs a slightly less folksy test.

Create an Original Twin for every edited image

An Original Twin is a structured record that permanently pairs a public image with the source image from which it came. It does not duplicate the entire marketing library into an AI prompt. It stores the pointers, fingerprints, edit classification, rules, approvals, and publication results needed to reconstruct what happened.

The record should exist before an altered image reaches a listing system, portal, email, text, brochure, social post, or property website.

A hash is useful. It is not magic. It proves that a file changed, not that the kitchen exists.

The Coalition for Content Provenance and Authenticity offers a stronger transport layer. C2PA Content Credentials can attach signed information about creation, tools, editing history, and ingredient assets to media. The credential is tamper-evident because cryptographic hashes bind the asset to its provenance record.8

C2PA also says manifests can be separated from the file. Durable credentials can use watermarking or fingerprinting to help rediscover them.8 The consumer project calls Content Credentials a “nutrition label for digital content.”9

Use that label when available. Keep the Original Twin anyway. Platforms resize, recompress, rename, and strip metadata with the casual efficiency of a teenager clearing a kitchen counter.

Build three Oppies around the image, not one inside the editor

Oppy supports custom AI employees across phone, email, SMS, chat, and more than 70 integrated business tools. Its current functions include image analysis, OCR, image generation, knowledge search, notes, checklists, conversation status, private administrator alerts, contact and CRM operations, scheduling, and workflow connections.10 11

That is enough to operate the control plane around property media. It is not a reason to let one agent generate, approve, and publish its own work.

The design is intentionally unglamorous. A good compliance system often looks like a librarian with excellent API access.

Give the Edit Classifier a strict prompt

SYSTEM PROMPT: ORIGINAL TWIN IMAGE CONTROL

You are Edit Classifier Oppy. Your job is to compare one governed
original property image with one proposed derived image and return a
structured classification for human review.

You do not publish images. You do not approve your own work. You do not
assume that disclosure makes a prohibited edit acceptable.

For every image pair:

1. Confirm that property_ref, original_url, original_hash, derived_url,
   destination, jurisdiction, and policy_version are present.
2. Compare visible elements in the original and derived images.
3. List every added, removed, obscured, relocated, recolored, resized,
   or structurally changed property element.
4. Classify the edit as one of:
   ORDINARY_PHOTO_CORRECTION,
   VIRTUAL_STAGING_PERSONAL_PROPERTY,
   MATERIAL_PROPERTY_ALTERATION,
   AI_GENERATED_PROPERTY_VIEW,
   or UNDETERMINED.
5. Treat furniture and decor separately from permanent or structural
   elements. Name both explicitly.
6. Retrieve the approved rule for the stated destination. Do not use a
   generic national rule when a destination-specific rule exists.
7. Set status to PASS only when the edit and required disclosure are
   allowed by the approved rule.
8. Set status to REVIEW when the change is ambiguous, the rule conflicts
   with another rule, or the property condition cannot be determined.
9. Set status to BLOCK when the original is missing, a permanent feature
   is fabricated or concealed contrary to policy, or the required original
   cannot accompany the derived image.
10. Return the disclosure text, required original-image placement, public
    original link requirement, and named human approver.

Never infer legal permission. When the rule is missing, route to the
responsible broker or counsel.

Then require a machine-readable answer.

{
  "twin_id": "twin_7f21",
  "property_ref": "listing_4812",
  "destination": "CRMLS",
  "policy_version": "CRMLS_11.5.2_2026-07-24",
  "edit_class": "VIRTUAL_STAGING_PERSONAL_PROPERTY",
  "changed_elements": [
    {"element": "sofa", "change": "added", "type": "personal_property"},
    {"element": "rug", "change": "added", "type": "personal_property"}
  ],
  "blocked_elements": [],
  "status": "PASS_WITH_DISCLOSURE",
  "disclosure": "Virtually staged. Original image follows.",
  "original_placement": "IMMEDIATELY_AFTER",
  "public_original_link_required": true,
  "human_gate": {"role": "responsible_broker", "decision": "PENDING"}
}

Your attorney, responsible broker, MLS, and compliance owner should approve the rule library and escalation design. A prompt can apply policy. It cannot enact one.

Test the ugly cases first

The staged living room is the easy case. Start quality assurance with the images most likely to produce a confident mistake.

The final row matters. Not every useful visualization belongs in public advertising. A buyer’s agent may use a concept image to help a client imagine a renovation, while a listing advertisement must accurately represent the property and disclose material alteration.7

The workflow should understand purpose before it admires pixels.

Make publication a reversible transaction

A Channel Publisher Oppy should prepare a release package, not fling an image toward the internet.

The package contains the original, the altered image, the Original Twin, the approved disclosure, destination ordering, the responsible human’s decision, and a rollback instruction. After publication, the Oppy reopens the public URL and checks that the label is visible, the original resolves, the pair is ordered correctly, and the image was not transformed into a new unreviewed derivative.

If any check fails, it removes or pauses the asset where the integration permits, sets the conversation to needs_attention, and sends a private administrator note. Oppy already documents that escalation pattern for sensitive work.11

The release is complete only when the public version passes.

The pattern travels beyond brokerage marketing

The same design applies wherever AI improves an image that affects a consumer, claim, asset, diagnosis, or legal position. The rule library changes. The need for an original does not.

Do not copy the real-estate rule into dental, insurance, auto, or legal work. Copy the architecture: source preservation, classified transformation, bounded use, clear disclosure, accountable approval, and public verification.

Measure whether the truth survives distribution

Count more than images generated. Track original-capture rate, unmatched derivatives, blocked permanent-feature changes, ambiguous edits routed to humans, original-link uptime, destination-policy freshness, post-publication failures, removal time, and repeat violations by source workflow.

A useful target is brutally simple: 100% of altered public property images have a retrievable original and a destination-valid disclosure.

The New York proposal reported on August 24 makes the emerging expectation unusually concrete. Its sponsor wrote that buyers and renters deserve the property’s current condition, “not simply its potential.”12

AI is very good at showing potential.

Your operating system must preserve condition.

References

[1] New York State Assembly, A11635 summary, actions, and bill text

[2] California Legislative Information, AB 723: Real estate digitally altered images disclosure

[3] RE Technology Inc. on X, post about AI, RESO, and real-estate data standards, August 24, 2026

[4] Boston Consulting Group, AI Hasn’t Fully Paid Off Yet in Logistics. But It Will., August 24, 2026

[5] CRMLS, Digitally Altered Image Guidance and FAQs, updated July 24, 2026

[6] NorthstarMLS, Guidelines for Virtual Staging and AI-Enhanced Listing Photos, July 10, 2026

[7] National Association of REALTORS®, Using AI to Enhance Listing Photos Can Be Legally Risky

[8] Coalition for Content Provenance and Authenticity, C2PA FAQ

[9] Content Credentials, About Content Credentials

[10] Oppy, AI workforce platform and agent capabilities

[11] Oppy, Oppy Functions: tools, permissions, and Needs Attention workflow

[12] The Post-Journal, Limits on AI in Real Estate Photos Introduced, August 24, 2026