Microsoft’s Copilot has quietly gained the ability to accept multiple files and images in a single chat session — a practical, long-requested update that promises to speed workflows and make multimodal reasoning more useful — but early hands‑on tests expose important limits and some surprising blind spots, particularly when the assistant is asked to identify specific graphics‑card models from images.
Microsoft has been steadily folding more capable multimodal features into Copilot across Windows, Edge, Microsoft 365, and Copilot Studio. The recent rollout adds a multi‑file upload UX to the consumer Copilot experience: users can attach more than one asset to the conversation and ask the assistant to “read across” them — compare, contrast, summarize or spot differences — instead of treating each file as a separate one‑off input. This builds on earlier Copilot improvements such as Vision sessions, semantic file search on Copilot+ devices, and the broader integration of the GPT‑5 family into Copilot’s reasoning stack.
At the same time, Microsoft’s documentation and product pages show that Copilot’s file capabilities are fragmented across products and tiers: Copilot (consumer) supports direct file uploads with explicit retention and file‑type rules, Copilot Studio and Microsoft 365 provide larger and more structured knowledge ingestion workflows, and OneDrive’s Copilot integration offers an independent “compare files” tool that accepts up to five files. Those product differences matter for how users should think about limits, privacy, and what “multi‑file” actually means in practice.
However, the PC Guide GPU experiment is a timely reminder that usefulness does not equal reliability. Copilot can reason across files and spot broad differences, but it is still prone to confidently producing precise, wrong labels when asked to perform fine‑grained product identification from images. That limit has immediate operational implications for tech teams, journalists, and asset managers who might otherwise lean on Copilot to automate verification tasks.
For everyday users, the feature will be a real productivity boost — provided they use it for ideation, summarization, or high‑level comparisons and treat any discrete, mission‑critical identification as a task that still needs human confirmation. For enterprises, the path forward should be cautious: pilot, validate, and gate by policy until Microsoft ships clearer enterprise controls, consistent quotas, and stronger provenance/validation tooling across Copilot surfaces.
Source: PC Guide Microsoft upgrades Copilot with new multi-file upload feature, so we tested its knowledge of GPUs
Microsoft has been steadily folding more capable multimodal features into Copilot across Windows, Edge, Microsoft 365, and Copilot Studio. The recent rollout adds a multi‑file upload UX to the consumer Copilot experience: users can attach more than one asset to the conversation and ask the assistant to “read across” them — compare, contrast, summarize or spot differences — instead of treating each file as a separate one‑off input. This builds on earlier Copilot improvements such as Vision sessions, semantic file search on Copilot+ devices, and the broader integration of the GPT‑5 family into Copilot’s reasoning stack.At the same time, Microsoft’s documentation and product pages show that Copilot’s file capabilities are fragmented across products and tiers: Copilot (consumer) supports direct file uploads with explicit retention and file‑type rules, Copilot Studio and Microsoft 365 provide larger and more structured knowledge ingestion workflows, and OneDrive’s Copilot integration offers an independent “compare files” tool that accepts up to five files. Those product differences matter for how users should think about limits, privacy, and what “multi‑file” actually means in practice.
What Microsoft shipped (quick realities)
- The Copilot conversation composer now allows attaching multiple files/images to a single chat so Copilot can reason across them rather than processing each in isolation. Early reporting and the PC Guide hands‑on note that multi‑file uploads are available in-app and in the browser for many users.
- Microsoft surfaces a set of supported file types for uploads (documents, common image formats and plain text formats); consumer Copilot docs list per‑file size/retention guidance such as a 50 MB per‑file guideline and automatic deletion windows for uploaded content. Product lines differ: Copilot Studio and Azure/Dataverse knowledge services accept much larger files (up to 512 MB in specific contexts) and allow grouping for agent knowledge.
- The multi‑file feature seems aimed at everyday productivity use cases: comparing job offers, building moodboards, studying with several notes, or asking Copilot to identify differences between versions of a document or photo set. OneDrive Copilot’s file‑compare takes this further by explicitly supporting comparing up to five OneDrive files.
The PC Guide test: GPUs, images and a cautionary tale
PC Guide’s practical test — uploading multiple photos of reviewed graphics cards and asking Copilot to spot the odd one out — demonstrates both the promise and the current weakness of image reasoning in Copilot.- The assistant managed to separate Founders Edition (FE) cards from partner/board partner cards, which is a useful baseline recognition capability. However, it misidentified the particular FE card in the image: Copilot labeled an RTX 5080 Founders Edition picture as an RTX 3090, and it named an ROG Astral card as an ROG Strix. Those are not minor typos; they reflect an inability to reliably read distinguishing design cues and model‑specific markings in some real‑world photos.
- The behavior was consistent across both quick and “Smart (GPT‑5)” response modes in the test, suggesting that (a) vision + reasoning pipelines are still imperfect when specific hardware model recognition is required, and (b) the error is not merely a latency/quality toggle in the UI but a gap in how visual evidence is mapped to precise product IDs.
Technical context: why the assistant confuses hardware details
There are several overlapping technical reasons an LLM+vision system will misidentify a particular hardware image:- Visual model limitations: object detectors and image‑understanding models are excellent at generic object classes (GPU, dog, car) but not always tuned for fine‑grained sub‑categories (RTX 4080 vs RTX 3090 vs RTX 5080) unless those classes have abundant, labelled training examples in the exact photographic conditions supplied. Off‑the‑shelf vision models often struggle with subtle badge/label differences or reflections on metallic surfaces.
- Multimodal grounding: success depends on how well the vision output is converted into symbolic facts the language model can reason with. If the vision stage is uncertain or returns noisy text/features, the language model may “fill in” plausible but incorrect labels.
- Training and updates: the assistant’s visual knowledge reflects the data it was trained on plus any document/image knowledge Microsoft has explicitly added to an agent or product. New card models (or vendor‑branded variants) may appear after training; the model can therefore default to older or more common product names.
- UI / model selection complexity: Copilot surfaces different “modes” (quick vs smart) and routes across model families (e.g., GPT‑5 family), but the endpoint’s decision to escalate to a deeper visual reasoning path is not a guarantee of correctness — only of more compute/attention on the prompt.
Strengths: why multi‑file upload is a meaningful upgrade
- Faster, more natural workflows: attaching several related files at once and asking a single question is how people actually work — reviewing contracts, comparing offer letters, or curating images for design briefs. Copilot’s multi‑file capability shortens that interaction loop and keeps context unified.
- Better multimodal reasoning: the system can now connect dots between files — for example, extract a table from one document, cross‑reference a graph from another, and summarize both — which is a different cognitive skill than handling isolated files. That’s particularly powerful for knowledge‑worker tasks and research synthesis.
- On‑device privacy and speed (Copilot+): when run on Copilot+ certified hardware with an NPU, semantic indexing and some inference can happen locally, reducing latency and surface privacy exposure for query processing. This on‑device-first approach is a meaningful design decision for privacy‑sensitive environments.
- Integration breadth: Copilot Studio and OneDrive offer complementary ways to work with multiple files (Copilot Studio for knowledge ingestion into agents, OneDrive’s compare for side‑by‑side document comparison). That diversification means Microsoft is enabling multi‑file scenarios across both consumer and enterprise surfaces. (microsoft.com, support.microsoft.com)
Risks, pitfalls and the governance questions you must ask
- Model accuracy vs. user trust: as the PC Guide GPU test shows, Copilot can confidently present incorrect identifications. When outputs carry operational consequences (asset inventory, procurement, legal review), false confidence is worse than silence.
- Confusing limits and inconsistent UX: file limits, daily upload caps and retention windows differ by Copilot flavor. Consumer Copilot pages list 50 MB per file and a 30‑day retention policy for uploaded files, while Copilot Studio supports much larger files for knowledge ingestion (up to 512 MB in some contexts) and OneDrive compare accepts five files. Users and admins must verify the exact product surface and subscription tier they’re using. (support.microsoft.com, learn.microsoft.com)
- Hardware gating creates a two‑tier Windows experience: Copilot+ NPUs unlock richer on‑device semantics but only on qualifying machines. That fragments expectations for users on older hardware. Enterprises must plan pilots carefully and avoid blanket enablement without testing.
- Privacy and compliance: Copilot surfaces recent files and will upload anything you explicitly attach; but the line between “surfacing” and “processing” must be understood by administrators. Verify whether indexes are stored locally, whether artifacts are backed up to cloud services, and how telemetry is handled under your organization’s compliance rules. Microsoft’s product notes emphasize explicit upload consent, but gaps and ambiguous retention behaviors have been flagged by testers.
- Operational limits and throttles: some enterprise users report daily upload limits or file‑count constraints in Copilot Chat or Copilot for Microsoft 365. Those product‑level throttles can break workflows that assume unlimited ingestion. Confirm service quotas before automating any high‑volume process.
Recommendations: how to use multi‑file Copilot safely and effectively
- Confirm which Copilot you’re using
- Consumer Copilot (taskbar / web) has different limits and a different retention policy than Microsoft 365 Copilot or Copilot Studio. Read the product’s file upload support page and your tenant’s admin notices before uploading sensitive material. (support.microsoft.com, learn.microsoft.com)
- Start with non‑sensitive test sets
- Before handing Copilot production artifacts, run a pilot with innocuous files: mock contracts, sample invoices, or photos of hardware. This helps calibrate how well Copilot extracts the exact facts you need and reveals hallucination patterns.
- Validate outputs with deterministic tools
- For tasks like inventory reconciliation or hardware identification, combine Copilot’s summary with deterministic checks: serial lookup services, GPU‑Z output, verified SKU metadata in inventory systems, or image‑based reverse search engines.
- Use the right surface for the job
- For long‑term, structured knowledge ingestion, prefer Copilot Studio’s grouped knowledge sources and Dataverse storage; for quick on‑the‑fly comparisons, OneDrive’s Compare Files tool may be the more predictable option. (learn.microsoft.com, support.microsoft.com)
- Audit retention and consent
- Treat every upload as a sharing event. Confirm file retention windows, encryption, and whether the product’s policy excludes training on uploaded content. Microsoft’s consumer Copilot docs state uploaded files are stored for a window (e.g., 30 days) and not used for training, but admin verification is prudent.
- Build guardrails for enterprise deployment
- Use Intune/MDM to control which devices and groups can use Copilot file features. Test DLP interactions, ensure Copilot telemetry is compatible with enterprise monitoring, and develop a rollback plan if unexpected data flows occur.
Where Microsoft should focus next (product critique)
- Improve fine‑grained vision tuning. The GPU misidentification underscores the need for specialized visual models or curated training sets for high‑value product classes (GPUs, phones, IoT SKUs) where exactness matters.
- Clarify limits and unify messaging across Copilot surfaces. Users are confused by differing file size caps, retention rules, and file‑count throttles across consumer Copilot, Microsoft 365, OneDrive, and Copilot Studio. A consolidated “file handling” matrix per product would reduce friction.
- Surface confidence and provenance. When Copilot identifies a specific model or extracts a claim from a file, show confidence scores and the visual evidence clipped from the image (highlighted badge/region) so users can verify quickly.
- Enterprise telemetry and retention controls. Provide clear admin controls for whether semantic indices are uploaded, encrypted backups of indices, and retention/erasure policies for Copilot+ devices. Enterprises will not adopt features they cannot audit or control.
- User correction flows. Allow users to flag and correct misidentified items and push those corrections into a tenant‑level model adapter so organizational knowledge improves over time without exposing data to model training.
Final assessment
Multi‑file upload in Copilot is a practical and user‑focused feature that brings Copilot closer to real productivity workflows: comparing documents, building visual moodboards, or asking cross‑file questions without tedious manual context switching. The update is a clear win for usability and for the concept of a single multimodal workspace.However, the PC Guide GPU experiment is a timely reminder that usefulness does not equal reliability. Copilot can reason across files and spot broad differences, but it is still prone to confidently producing precise, wrong labels when asked to perform fine‑grained product identification from images. That limit has immediate operational implications for tech teams, journalists, and asset managers who might otherwise lean on Copilot to automate verification tasks.
For everyday users, the feature will be a real productivity boost — provided they use it for ideation, summarization, or high‑level comparisons and treat any discrete, mission‑critical identification as a task that still needs human confirmation. For enterprises, the path forward should be cautious: pilot, validate, and gate by policy until Microsoft ships clearer enterprise controls, consistent quotas, and stronger provenance/validation tooling across Copilot surfaces.
Quick reference: five things to know right now
- Copilot can accept multiple files/images in one chat and reason across them, which shortens multi‑document workflows.
- Product limits vary: consumer Copilot pages suggest a per‑file guideline (e.g., 50 MB) and a retention window; Copilot Studio supports larger, grouped uploads for agent knowledge (up to 512 MB in contexts). (support.microsoft.com, learn.microsoft.com)
- OneDrive Copilot’s Compare Files feature explicitly supports up to five files for side‑by‑side analysis.
- Visual recognition can separate broad classes (Founders Edition vs partner card) but may misidentify exact models (RTX 5080 FE labeled RTX 3090 in PC Guide’s test), so verify critical calls.
- If deploying in an organization, pilot the feature, verify retention and telemetry, and use Intune/MDM policies to control who can attach files to Copilot chats.
Source: PC Guide Microsoft upgrades Copilot with new multi-file upload feature, so we tested its knowledge of GPUs